Changelog

All notable changes to BlockRun, newest first — gateway endpoints, model lineup, pricing, and SDK releases.

[2026-11-04]

Changed — Grok Imagine Pro generates at medium quality, at a new price

• xAI retired the model behind xai/grok-imagine-image-pro on 2026-11-02 and began serving its requests at the low quality of Grok Imagine 2.0. Pro now requests Grok Imagine 2.0 at medium quality, one tier above xai/grok-imagine-image-2.0.
• A 1024x1024 image is quoted at $0.064, up from $0.0535. The 402 quote always shows the current price before you pay.
• For the lower price, use xai/grok-imagine-image-2.0 at $0.043 per image. The model id xai/grok-imagine-image-pro is unchanged.

[2026-11-03]

Changed — Pay-per-call chat is charged at your full max_tokens

• The 402 quotes input plus output at your max_tokens. An exact payment settles that quote, so set max_tokens to what you need. Pay with upto (Base) or batch settlement (Solana) and you are charged only the tokens produced.
• On blockrun.ai an exact chat payment used to be quoted at a share of max_tokens as output. It is now quoted, and settled, at the full max_tokens, the same as exact on sol.blockrun.ai. A request with no max_tokens is quoted at the default 8192 (or the model's maximum output, if lower). Example: openai/gpt-5.5 with a one-word prompt and no max_tokens is now quoted $0.246850; with "max_tokens": 1000 it is $0.031090.
• upto is offered on Base as a second accepts entry on the models that support it, currently the DeepSeek and Anthropic models. It authorizes the same ceiling and settles only the tokens produced. Select the entry by scheme, not by position.
• The GPT-6 reasoning-effort uplift from 2026-10-10 no longer changes a price: the quote is already the full ceiling.
• This applies to /v1/chat/completions, /v1/messages and /v1/responses (where the ceiling is max_output_tokens, and an unset one is the model's maximum output, not 8192). Metered API-key billing (user.blockrun.ai, api.blockrun.ai) bills actual usage and is unaffected. See How the quote is computed.

Changed — Async media is released once its payment confirms

• On image generation, image editing, music and video, a finished job's result is now handed over only after its payment is confirmed on-chain. While the transfer confirms, the poll (or the POST, on the fast path) answers 202 with payment_status: "settling" and Retry-After: 3. Keep polling poll_url with the same wallet; no second charge is taken while a transfer is unresolved.
• A transfer that confirms under a different transaction hash, for example after the facilitator bumped its gas, is still recognised as paid.
• Cancelling the signed authorization while its transfer is pending counts as a failed payment: the result is not released.

Changed — Errors never carry upstream text on a 429 or 5xx

• debug now appears only on a caller-fault 4xx, redacted. Rate limits, upstream failures and gateway errors return just code and message. Quote the x-blockrun-gateway-request-id response header when you report one. See Errors.

Added — Model-swap headers on streams, readable from browsers

• A paid stream that fails over to its fallback model now sends X-Fallback-Used, X-Original-Model and X-Fallback-Model, as non-streaming responses already did.
• Model-swap headers are listed in Access-Control-Expose-Headers whenever they are present.
• Swap headers and response bodies name public model ids. When the same model answers from another endpoint, it is no longer reported as a swap.

Changed — Errors inside a stream carry a code, not upstream text

• A failure after a stream has started ends with a fixed message and a code: deadline, stall, rate_limited or upstream_error. An upstream rejection of one of your parameters still gives its redacted reason. See Errors.

Fixed

• Unpaid requests with an empty or invalid body (discovery crawlers, uptime monitors) got a 500 from 2026-11-02 03
  UTC to 2026-11-03 00
  UTC instead of the 402 challenge. Paid calls were unaffected.
• sol.blockrun.ai now redirects http:// to https:// with a 301 and sends Strict-Transport-Security, matching blockrun.ai.
• llms.txt no longer lists the internal-only /v1/zerox/trade-analytics/* endpoints.
• Prediction-market paths with a template placeholder instead of an id (for example REPLACE_WITH_MARKET_ID, TOKEN_ID, test) now return 404 before any payment is requested, instead of being charged.

[2026-10-25]

Added — Kimi K3.1 Pro

• moonshot/kimi-k3.1-pro ($2.00/M in · $10.00/M out, $0.20/M cached, 1M context, image + text input, returns reasoning_content) — Moonshot's post-trained successor to Kimi K3, built for long-horizon coding and agents. Cheaper than K3's $3.00/$15.00. Verified live with real completions, streaming, forced and automatic tool calls, and vision before listing.
• Kimi K3.1 Pro replaces K3 as the featured Kimi. moonshot/kimi-k3 is hidden from the public list but stays routable.
• Visible model counts are unchanged.

[2026-10-10]

Changed — GPT-6 calls are priced by reasoning effort (superseded 2026-11-03: exact chat is now quoted at the full max_tokens)

• A pay-per-call chat request is charged before it runs: its input plus a tenth of its max_tokens as output. On GPT-6, higher reasoning_effort spends more reasoning, which bills as output; measured on GPT-6.1 Sol, high ran up to 2.3x the output of medium, xhigh up to 2.5x, max up to 4.4x.
• On every GPT-6 model the output share of the quote now scales with the effort requested: 2.5x at high, 3x at xhigh, 4.5x at max, never past max_tokens. No effort, low and medium are charged exactly as before. On plain chat, where max is served as xhigh, it is charged as xhigh.
• This applies to /v1/chat/completions and /v1/responses. Metered API-key billing (user.blockrun.ai, api.blockrun.ai) bills actual usage and is unaffected.

[2026-10-09]

Changed — GPT-6 reasoning effort

• reasoning_effort: "max" now works on every GPT-6 model. It used to fail with a 400 on chat completions. On plain chat the gateway now lowers it to xhigh; on tool calls, which the gateway serves through OpenAI's Responses API, it is sent as max.
• GPT-6 tool calls now run at the effort you ask for. They used to run at medium whatever the request said, so a low call billed medium reasoning; now it bills low. With no reasoning_effort, tool calls still default to medium.
• none and minimal are raised to low on gpt-6-astra and gpt-6.1-sol. gpt-6-sol and gpt-6-luna accept none; minimal is raised to low on all four.

Added — GPT-6.1 Sol

• openai/gpt-6.1-sol ($2.00/M in · $10.00/M out, 1M context, 128K max output, vision, reasoning) — OpenAI's 2026-09-29 refresh of the GPT-6 Sol tier, knowledge cutoff 2026-04-30. Same price as gpt-6-sol, but cached input is $0.10/M, half of Sol's $0.20/M; cache writes are $2.50/M.
• Like the rest of GPT-6, it reprices the whole request at 2x input and cached rates, 1.5x output once the prompt exceeds 272K tokens — $4.00/$15.00 above the threshold — as OpenAI bills it.
• Tool calls work: the gateway routes them through OpenAI's Responses API, so no client change is needed. reasoning_effort accepts low through max (see the effort change above); none and minimal are raised to low. No Flex tier yet.
• Falls back to gpt-6-sol, at the same rate, if its upstream fails.
• Visible chat models are now 88; total catalog 115.

[2026-10-07]

Changed — Decisions replaces OpenJev; free with a key, on api.blockrun.ai only

• Typed judgments are now OpenAI Decisions-compatible (gpt-6-luna): predicate, choice and score questions over text or images. POST https://api.blockrun.ai/v1/decisions, free for any registered key — register at user.blockrun.ai and create one. Any OpenAI SDK works with base_url="https://api.blockrun.ai/v1".
• Not sold over x402. There is no Decisions route or price on blockrun.ai: the x402 minimum payment is above what a judgment costs.
• OpenJev is retired. POST https://api.blockrun.ai/v1/decide stays as a compatibility alias that translates the old shape; see Decide (legacy).

Added — Mistral Large 4

• mistral/mistral-large-4 ($0.68/M in · $2.09/M out, 512K context, 262K max output) — Mistral's flagship, released 2026-10-06. Reasoning, tool calling, JSON mode and image input. It reasons by default and the reasoning bills as output, so leave room in max_tokens.
• The price is Mistral's launch rate, half its $1.36 / $4.18 list. When Mistral ends the promotion the price here moves with it.
• Visible chat models are now 87; total catalog 114.

[2026-10-02]

Removed — Gemma 4 31B (free), the same day

• nvidia/gemma-4-31b is delisted. The free upstream stopped answering: no response inside 120 seconds, on short and realistic prompts alike. Calls now redirect to nvidia/nemotron-3-super-120b, and it comes back when the upstream does.
• Visible chat models are now 86; total catalog 113; free models 7.

Added — Qwen3.8 Max, Tencent Hy4 Preview, Laguna S 2.1, Muse Glimmer 30B, Gemma 4 31B on Base

• qwen/qwen3.8-max ($2.00/M in · $6.00/M out, 1M context, 131K max output) — the newest Qwen flagship. Falls back to Qwen3.7 Max, which costs less, if its upstream fails.
• tencent/hy4-preview ($0.834/M · $2.501/M, 1M context, 64K max output) — Tencent's 770B open-weight MoE. It reasons before answering; a small max_tokens can be spent entirely on reasoning.
• poolside/laguna-s-2.1 ($0.09/M · $0.18/M, 1M context, 131K max output) — Poolside's 118B open-weight coding model, the paid sibling of the free Laguna XS 2.1.
• nvidia/muse-glimmer-30b (free, 131K) — Meta's open 30B after Llama. Reasons first; ask for 1000+ max_tokens.
• nvidia/gemma-4-31b (free, 131K) — Google's Gemma 4.
• Solana (sol.blockrun.ai) has listed all five since 2026-09-07; Base (blockrun.ai) now matches. Each was verified with real completions on the upstream the gateway uses, and each paid price is the upstream's rate.
• Visible chat models are now 87; total catalog 114; free models 8.

[2026-09-29]

Added — Claude Opus 5.5, Claude Sonnet 5.5, Grok 4.7

• anthropic/claude-opus-5.5 ($4.00/M in · $20.00/M out, 1M context, 128K max output, vision) — the newest Opus, and cheaper than Opus 5 ($5/$25). Cache reads are $0.20/M, 0.05x input rather than the usual 0.1x. Thinking is always on; set a low output_config.effort to keep it short.
• anthropic/claude-sonnet-5.5 ($2.00/M · $10.00/M, 1M context, 128K max output, vision) — the newest Sonnet at Sonnet 5's price. Falls back to Sonnet 5 at the same rate if its upstream fails.
• xai/grok-4.7 ($2.00/M · $6.00/M, 500K context, vision, reasoning effort low–xhigh) — xAI's newest flagship. Doubles to $4.00/$12.00 once a prompt reaches 200K tokens, as xAI bills it. Falls back to Grok 4.6 at the same rate.
• Opus 5.5 and Sonnet 5.5 reject a forced tool choice and a {"type": "disabled"} thinking block. The gateway serves a forced choice as auto, drops disabled on Opus 5.5 (it cannot turn thinking off), and translates it to Sonnet 5.5's own off switch, between_tools. Requests written for Opus 5 / Sonnet 5 work unchanged.
• All three were verified with real completions on the production upstreams before listing, available on both Base (blockrun.ai) and Solana (sol.blockrun.ai).
• Visible chat models are now 82; total catalog 109.

[2026-09-25]

Added — GPT-6 Sol and GPT-6 Luna

• openai/gpt-6-sol ($2.00/M in · $10.00/M out, cached $0.20/M, 1M context, 128K max output, vision) — the cost-efficient GPT-6 tier below the Astra flagship, for complex coding and agentic work.
• openai/gpt-6-luna ($0.10/M · $0.50/M, cached $0.01/M, 1M context, 128K max output, vision) — the smallest GPT-6 tier.
• Like gpt-6-astra, both reprice the whole request at 2x input and cached rates, 1.5x output once the prompt exceeds 272K tokens — gpt-6-sol is $4.00/$15.00 above the threshold — as OpenAI bills it.
• Unlike Astra, both accept reasoning_effort: "none". Both reject temperature other than 1, top_p, penalties, logprobs and stop; reasoning_effort: "max" is rejected upstream despite the model page listing it (since 2026-10-09 the gateway accepts it; see that entry).
• Both serve Flex (service_tier: "flex", half the standard rate), verified with a paid call on 2026-09-26 against a no-tier control.

[2026-09-08]

Added — Claude Fable 5.1, Grok 4.6

• anthropic/claude-fable-5.1 ($10.00/M · $50.00/M, cached $0.25/M, 1M context, 128K max output, vision) — Anthropic's most capable model, successor to Fable 5 in the same tier at the same price. It rejects a forced tool choice; the gateway serves tool_choice: "required" or a named tool as auto. Anthropic's dashed id claude-fable-5-1 resolves too.
• xai/grok-4.6 ($2.00/M · $6.00/M, cached $0.50/M, 500K context, vision, reasoning effort) — doubles to $4.00/$12.00 once a prompt reaches 200K tokens.

[2026-09-07]

Added — GPT-5.1

• openai/gpt-5.1 ($1.25/M · $10.00/M, 400K context, 128K max output, vision and tools).

Fixed — long-context tier removed where OpenAI does not charge one

• openai/gpt-5.2, gpt-5.2-pro, gpt-5.3-codex, gpt-5.4-mini and gpt-5.4-nano were repricing requests above 272K prompt tokens; OpenAI bills those five flat. They are flat now.

[2026-09-05]

Added — GPT-6 Astra

• openai/gpt-6-astra ($10.00/M · $50.00/M, cached $1.00/M, 1M context, 128K max output, vision) — OpenAI's GPT-6 flagship for long-horizon agentic work and computer use. 2x input / 1.5x output above 272K prompt tokens ($20.00/$75.00). It rejects reasoning_effort: "none".

Fixed — GPT-5.6 Sol at OpenAI's cut price

• openai/gpt-5.6-sol and openai/gpt-5.6-sol-pro bill $4.00/$20.00, following OpenAI's 2026-08-21 cut from $5.00/$30.00.

[2026-08-30]

Added — Qwen3.8 Flash, DeepSeek V4 Flash Vision, MiMo-V2.5

• qwen/qwen3.8-flash ($0.15/M in · $0.47/M out, 1M context, image input) — 125B MoE with GDN+QSA hybrid attention. Alibaba positions it above Qwen3.7 Plus, which we sell at $0.32/$1.28, so it is both newer and cheaper than the tier it beats.
• deepseek/deepseek-v4-flash-vision-exp ($0.44/M · $1.32/M, 1M context, image input) — our first DeepSeek model that takes images. Priced at DeepSeek's peak rate, since their new peak/off-peak split (peak 01
  –04
  and 06
  –10
  UTC, Mon–Fri) would otherwise put it below cost for seven hours every weekday.
• xiaomi/mimo-v2.5 ($0.14/M · $0.28/M, 1M context, image input) — the natively multimodal MiMo SKU, distinct from (and a third the price of) the text-only mimo-v2.5-pro we already listed.
• All three were probe-verified with real completions — and real image input — before listing.

Fixed — two GPT-5.6 Pro SKUs were unroutable, not just mispriced

• OpenAI raised openai/gpt-5.6-terra-pro to $2.00/$12.00 and openai/gpt-5.6-luna-pro to $0.20/$1.20. Our catalog still carried the old $1.00/$6.00 and $0.10/$0.60, and the gateway hands upstream a cost ceiling derived from the catalog price — so every call was rejected with "no endpoints satisfy the max price". Luna Pro (no fallback) returned a hard error; Terra Pro silently fell through to its cheaper fallback while billing Terra Pro's rate. Both are repriced to the current upstream rate.
• Visible chat models are now 76; total catalog 100.

Changed — free tier rebuilt after NVIDIA retired four of the five free models

• Delisted nvidia/step-3.7-flash, nvidia/nemotron-nano-9b-v2 and nvidia/nemotron-nano-12b-v2-vl (published 410 Gone on both passes of a live --real probe) and nvidia/mistral-nemotron (still listed upstream, but a completion never returns: >150s, zero bytes, both passes). Calls pinned to any of these ids are redirected to a healthy free model and still return 200.
• Also 410: the hidden nvidia/nemotron-super-49b, which was simultaneously the free cascade's tertiary rung and the fallback of its primary — both retargeted in the same change.
• Added nvidia/nemotron-3.5-lightning (thinking-mode reasoning, 131K context, ~35 tok/s), nvidia/nemotron-3-nano-30b (~121 tok/s, the fastest free model in the catalog) and nvidia/llama-3.2-11b-vision (Meta Llama 3.2, 128K context, image input) — each verified with a real completion through the gateway before listing.
• nvidia/gpt-oss-120b and nvidia/gpt-oss-20b recovered upstream and no longer redirect elsewhere.
• nvidia/nemotron-3.5-lightning and nvidia/nemotron-3-nano-omni-30b-a3b-reasoning now serve from a larger free-tier capacity pool, with their previous route kept as the fallback — same models, 4.9s median against 16.3s, and Lightning gains a 1M-token context (was 131K). Image input verified through the new route before moving the vision model.
• Added three more free models: nvidia/nemotron-3-ultra-550b (550B/55B MoE, 1M context — the largest free model in the catalog), cohere/north-mini-code (compact coding, sub-second) and poolside/laguna-xs-2.1 (coding, ~161 tok/s).
• Visible chat models are now 73 (was 71); total catalog 97; free models 7 (was 5).

Fixed — an upstream error delivered as HTTP 200 no longer reads as an empty answer

• OpenAI-compatible relays can answer 200 with an error object and no choices when the provider behind them fails ({"error":{"message":"Upstream error from Nvidia: Service temporarily overloaded","code":502}}). The gateway treated that as a successful empty completion, so the fallback chain, the free-tier health breaker and the retry classifier all saw a success — and a paid model still billed the minimum charge. Both the blocking and the streaming path now rethrow it as a real upstream error.

Fixed — 402 responses now carry the payment challenge in the body, not just the headers

• Every 402 Payment Required body now spreads x402Version and accepts at the top level, mirroring the signed challenge that has always lived in the PAYMENT-REQUIRED / X-Payment-Required / WWW-Authenticate headers. Pre-v2-era x402 clients (early x402-fetch/x402-axios, and some third-party wrappers) only ever parsed the body; finding no top-level accepts there, they silently gave up instead of auto-paying — invisible in our logs, indistinguishable from organic non-conversion.
• Applies across every paid endpoint — chat completions, responses, messages, images, video, music, speech, search, market data, RPC, Modal sandboxes, RealFace, Virtual Portrait, and Polymarket funding. A route's own fields still win on any key collision with the mirrored ones. (BlockRunAI/blockrun#446)

[2026-08-29]

Removed — OpenAI GPT-5.3

• Delisted openai/gpt-5.3 — OpenAI has deprecated it upstream. Requests that still name it are redirected to openai/gpt-5.2 (same $1.75/M in · $14.00/M out), so existing callers keep working; openai/gpt-5.3-codex is unaffected.
• Visible chat models are now 71 (was 72); total catalog 95 across chat, image, video, music, speech, and sound effects.

[2026-08-27]

Added — Z.AI GLM-5.3 Flash, the first natively multimodal GLM-5

• Added zai/glm-5.3-flash ($0.15/M in · $0.50/M out, 1M context, 131,072 max output, image input, always-on reasoning) — a 320B/18B hybrid-attention MoE that takes screenshots and rendered UI directly into the coding loop. Verified live on Z.AI before listing: real completions, tool calls, streaming, and vision.
• Priced at Z.AI's list rate. Z.AI is running a 50% launch promotion through 2026-09-09; we list the post-promotion price rather than one that would lapse under our cost two weeks later.
• Thinking cannot be disabled on the 5.3 generation, and does not need to be — unlike GLM-5/5.1/5.2, it returns content alongside its reasoning.
• zai/glm-5.3 ($1.40/$4.40, 1M context), launched 2026-08-19, is now listed in the model reference table as well.
• Visible chat models are now 72 (was 71); total catalog 96 across chat, image, video, music, speech, and sound effects.

Changed — Reasoning rides in reasoning_content, never as tool-call content

• /v1/chat/completions non-streaming responses now carry a model's chain of thought on choices[].message.reasoning_content, the same field streaming has always used for delta.reasoning_content. It is optional and absent when the model produced no reasoning. Claude thinking, DeepSeek, GLM, MiniMax, Kimi and open-weight reasoning models all populate it. (#423)
• An assistant turn that carries tool_calls now has content: "", as the OpenAI spec intends. Previously the gateway rescued the empty content with the model's raw reasoning, so Anthropic-format clients rendered the chain of thought as a visible text block ahead of the tool_use, and clients echoing history fed reasoning back as content. Turns with no tool calls and no answer still receive the reasoning text as content so a paid call never returns a blank string. (#422)
• /v1/messages builds a leading thinking block from the same field, so streaming and non-streaming Anthropic-format responses finally have the same shape.

[2026-08-26]

Changed — Payment-verification 402s carry a machine-readable code

• Every BlockRun-native paid endpoint now answers a failed PAYMENT-SIGNATURE with 402 and a code a client can branch on instead of parsing prose: PAYMENT_UNFUNDED (the on-chain simulation reverted — usually an empty wallet, or an authorization outside its validity window), PAYMENT_BLOCKHASH_STALE (a Solana-signed transaction pinned to a blockhash that expired — re-sign against a fresh one and resend; nothing was charged), PAYMENT_REPLAY (nonce already used), and PAYMENT_INVALID for everything else. Previously only /v1/chat/completions classified failures; the other paid routes returned a bare {error, details}.
• The facilitator's own explanation is now folded into debug rather than only logged, so a rejection says which failure it was. Verification always runs before settlement, so any of these codes means nothing was charged.
• The vendor-compatible /v1/messages, /v1/responses and Gemini-native endpoints keep their vendor error schemas and do not carry the field. (commit ac50282)

[2026-08-19]

Added — Z.AI GLM-5.3

• Added zai/glm-5.3 ($1.40/$4.40 per 1M, 1M context, 131,072 max output). Thinking is always on for this generation and cannot be disabled — unlike GLM-5/5.1/5.2 it returns content alongside its reasoning, so no gateway workaround is applied. (#397)

[2026-08-13]

Added — xAI Grok Imagine Video 1.5, and official resolution tiers on both Grok video SKUs

• Added xai/grok-imagine-video-1.5 (image/text-to-video, 480p/720p/1080p, up to 15s, native audio) at xAI's official per-second rates: $0.08/s @480p (default), $0.14/s @720p, $0.25/s @1080p. (#380)
• xai/grok-imagine-video moved to the same tiered model: $0.05/s @480p (default) and $0.07/s @720p; 1080p/4K are rejected with a 400 before payment. resolution is now a billing tier and is pinned in the quote, so the price you sign is the tier you get. (#378)
• Both Grok SKUs charge the official rate exactly (no 5% media margin) plus a flat $0.001 per generation, disclosed as perGenerationFee in the 402 price block alongside pricePerSecond for the billed tier; the flat $0.001 transaction fee applies on top. aspect_ratio (incl. 3:2 / 2:3) is now validated and forwarded — it was previously accepted and silently dropped. (#379, #381)

[2026-08-12]

Added — Seedance 2.0 Mini

• Added bytedance/seedance-2.0-mini at $0.0797/sec (480p/720p with audio, 4–15s) — roughly half the rate of seedance-2.0-fast for the same 720p output, the sensible default for volume. Supports the same image, first/last-frame and RealFace inputs as the rest of the 2.0 family. (#366, #369)

[2026-08-08]

Changed — Per-token chat now bills at list price, with no platform margin

• Every per-token chat model on /v1/chat/completions, /v1/messages and /v1/responses is now priced at the maker's list rate with 0% platform margin (was 5%). The flat $0.001 per-request transaction fee is unchanged and is the only thing BlockRun adds to a chat call. Image, video, music, speech and Live Search keep their 5% margin. (#354)
• Three prices were corrected in the same pass: deepseek/deepseek-chat and deepseek/deepseek-reasoner to $0.14/$0.28 (from $0.20/$0.40, per DeepSeek's current list), and zai/glm-5 to $1.00/$3.20 (from $0.60/$1.92 — Z.AI had raised its standard rate).
• Images: a client that loses the response to an inline image generation can now retry with the same signed authorization and get the job it already paid for back (402 / PAYMENT_REPLAY now carries job_id and poll_url) instead of being told to sign — and pay — again. (#346)

[2026-08-07]

Added — Seedance 2.5, and the Seedance family repriced

• Added bytedance/seedance-2.5 — 4–30s clips (double the 2.0 ceiling) at a 720p ceiling, $0.315/sec. It is not a straight upgrade over 2.0, which keeps the 4K ceiling at 4–15s; both stay listed. Duration always defaults to 5s (the model's own "-1, pick for me" default is never forwarded, since the price is signed before generation). (#351)
• The rest of the family moved to current per-second rates: seedance-1.5-pro $0.070/sec, seedance-2.0-fast $0.165/sec, seedance-2.0 $0.227/sec (720p, 5s default; 1080p and 4K scale by area). Per-model resolution and duration limits are now derived from the model's own parameter schema and enforced before the 402. (#353)

[2026-08-05]

Changed — Discovery now lists every live endpoint, with derived prices

• /.well-known/x402 and openapi.json now publish the full live surface — Surf, 0x, DefiLlama, the templated Predexon paths, every RPC chain, and the first-party phone, voice, music, RealFace, Portrait, Polymarket-funding, onramp and catalog endpoints — instead of a hand-maintained subset. Every published price is derived from the same constant the 402 signs, which corrected five stale literals (Predexon, phone buy/lookup/list, DefiLlama). (#336)
• Six dead prediction-market endpoints were de-registered (matching-markets, matching-markets/pairs, markets, listings, outcomes/{id}, dflow/*). (#332, #336)

[2026-08-03]

Changed — GPT-5.6 Luna is 80% cheaper, Terra 20%, passed through in full

• OpenAI cut GPT-5.6 pricing on 2026-07-30. Both cuts are now live on the gateway: openai/gpt-5.6-luna goes to $0.20/$1.20 per 1M (from $1.00/$6.00, -80%) and openai/gpt-5.6-terra to $2.00/$12.00 (from $2.50/$15.00, -20%). Long-context tiers follow at $0.40/$1.80 and $4.00/$18.00.
• openai/gpt-5.6-sol is unchanged at $5.00/$30.00 — OpenAI left the flagship alone. Luna now costs less than gpt-5.4-mini.
• We passed the cut through rather than keeping the spread. Every other OpenAI model in the catalog was re-checked against the published rates in the same pass and already matched.
• Fallbacks moved with the prices: Terra now falls back to gpt-5.4-mini and Luna to gpt-4o-mini. The previous targets were not repriced by OpenAI, so they now cost more than the primaries bill and would settle below cost on every failover.

Added — Eight models: Gemini 3.6 Flash, GPT-5.6 Pro tiers, Qwen3.7 Plus/Flash, Nano Banana 2

• google/gemini-3.6-flash ($1.50/$7.50 per 1M) and google/gemini-3.5-flash-lite ($0.30/$2.50). Thinking is always on for both — they reject a zero thinking budget upstream, which the gateway no longer sends.
• openai/gpt-5.6-sol-pro ($5.00/$30.00), openai/gpt-5.6-terra-pro ($1.00/$6.00), openai/gpt-5.6-luna-pro ($0.10/$0.60) — the pro reasoning modes behind ChatGPT Pro. Served on /v1/chat/completions; /v1/responses now gates on the OpenAI model family rather than on the id prefix.
• qwen/qwen3.7-plus ($0.32/$1.28, 131K max output) and qwen/qwen3.7-flash ($0.03/$0.13) complete the Qwen3.7 line; both are reasoning models.
• google/nano-banana-2 ($0.09/image, 1024×1024) on both /v1/images/generations and /v1/images/image2image. (#329)

Changed — RPC upstream faults are ours, not yours

• /v1/rpc/{network}: an upstream timeout, credential failure or 5xx now returns 502 with Retry-After: 30 and releases your payment nonce, so an SDK retry with the same signed header succeeds instead of dying as a false PAYMENT_REPLAY. A known chain's upstream 404 is reported as our stale slug, not a caller error. (#324, #327, #328)

[2026-08-02]

Changed — The 402 tells you what to fix

• /v1/chat/completions: an unfunded wallet now gets 402 / PAYMENT_UNFUNDED with a plain-language explanation instead of "message a human on Telegram". One empty wallet had retried 292 times against the old message. (#322)
• The chat 402 body's price.amount now equals the amount the header signs, transaction fee included. It had been quoting the pre-fee price, under-stating a floored call by the whole $0.001 fee. (#321)
• When the gateway serves a smaller max_tokens than you asked for (model ceiling or context headroom), the response now says so: X-Max-Tokens-Capped: true, X-Max-Tokens-Requested, X-Max-Tokens-Effective. Headers only, only when the clamp bit; bodies are unchanged. (#271)

[2026-07-29]

Changed — Transaction fee back to $0.001

• The flat per-request transaction fee added to every paid call returns to $0.001 (it had been $0.002 since 2026-07-11). The smallest possible paid chat call is therefore $0.002 all-in. Every published minimum was swept to match. (#319, #320)

[2026-07-24]

Added — Anthropic Claude Opus 5, live on launch day

• Added anthropic/claude-opus-5 ($5.00/$25.00 per 1M, 1M context, 128K max output, adaptive thinking, vision), Anthropic's newest flagship — a step-change over Opus 4.8 for deep reasoning and agentic coding at the same price. Verified live before listing.
• claude-opus-4.8 stays listed and routable at the same rates.
• Visible chat models are now 64 (was 61); total catalog 84 across chat, image, video, music, speech, and sound effects.

[2026-07-20]

Added — Qwen3.7 Max (Alibaba flagship)

• Added qwen/qwen3.7-max ($1.48/$4.43 per 1M, 1M context), Alibaba's Qwen flagship — the Max tier, strong reasoning + coding + agentic tool use. Verified live before listing. Fills the gap where the only Qwen models were the free tier.
• Visible chat models are now 61 (was 60); total catalog 81 across chat, image, video, music, speech, and sound effects.

[2026-07-17]

Added — Moonshot Kimi K3

• Added moonshot/kimi-k3 ($3.00/$15.00 per 1M, 1M context, image + text input, returns reasoning_content), Moonshot's new flagship — a 2.8-trillion-parameter open MoE. Verified live on release day with real completions, tool calls, and vision before listing.
• kimi-k3 replaces kimi-k2.7 as the featured Kimi. K2.7 is now hidden from the public list but remains routable, and new traffic auto-prefers K3 via fallbackModel chaining.

[2026-07-13]

Added — xAI Grok 4.5

• Added xai/grok-4.5 ($2.50/$9.00 per 1M, 500K context, vision + reasoning), xAI's flagship. Live Search supported (+$0.025/source). Verified with real completions on every serving route before listing.

Changed — Grok long-context pricing now mirrors xAI's official 2x tier

• xAI bills the Grok family at 2x above 200K prompt tokens. BlockRun now applies the same tier: once a request's prompt reaches 200K tokens, the whole request reprices at the long-context rate — grok-4.5 $5.00/$18.00, grok-4.3 $3.00/$8.00, grok-build-0.1 $3.00/$6.00, grok-4.20 $4.00/$12.00 per 1M. Below 200K, base rates are unchanged.
• grok-4.5 fallback set to grok-4.20-reasoning (Live-Search-capable) so a failover never silently drops search.

[2026-07-09]

Added — OpenAI GPT-5.6 family (Sol / Terra / Luna), live on GA day

• Added all three GPT-5.6 tiers the day OpenAI moved them to general availability: openai/gpt-5.6-sol ($5.00/$30.00 per 1M — flagship reasoning/coding/agentic), openai/gpt-5.6-terra ($2.50/$15.00 — balanced), openai/gpt-5.6-luna ($1.00/$6.00 — high-volume, latency-sensitive). 1M context each. All verified with real completions and tool calls before listing.
• gpt-5.6-sol is now the featured OpenAI flagship (replaces gpt-5.5 on the homepage table at the same price; gpt-5.5 remains available).
• Gateway compatibility handled transparently: GPT-5.6 rejects function tools on /v1/chat/completions unless reasoning_effort is explicitly "none" — BlockRun sets this automatically whenever tools are present, so agent frameworks work unchanged.

[2026-06-28]

Added — response_format: json_schema (OpenAI structured outputs)

• /v1/chat/completions now accepts response_format: {"type":"json_schema","json_schema":{…}} and forwards it verbatim. Schema-guaranteed output is enforced natively on OpenAI (GPT) models; other providers receive it on a best-effort passthrough. Previously only json_object/text were accepted (any other type returned a 400).
• The thinking block is also accepted more leniently: {"type":"disabled"} and blocks without budget_tokens no longer 400.

[2026-06-24]

Added — Real 4K on Seedance 2.0

• bytedance/seedance-2.0 now accepts resolution: "4K" and renders true 3840×2160 (re-verified with ffprobe). Billed at 9× the 720p token rate (area-proportional, same basis as 1080p=2.25×).
• 4K is seedance-2.0 only — seedance-1.5-pro and seedance-2.0-fast still cap at 1080p/1K and return 400 for 4K. 2K remains unavailable on all Seedance models (not produced upstream).

Changed — Seedance video max duration raised

• bytedance/seedance-2.0 and bytedance/seedance-2.0-fast: max duration 10s → 15s.
• bytedance/seedance-1.5-pro: max duration 10s → 12s.
• Default duration stays 5s; price scales linearly with duration (Seedance is token-metered, so a 15s clip costs 3× a 5s clip at the same resolution). Services page copy updated from "5–10s" to "5–15s".

[2026-06-22]

Added — Cache and reasoning tokens surfaced in usage fields

• Non-stream Claude (/v1/messages and OpenAI-compat) responses now emit cached_creation_tokens (cache-write count) under prompt_tokens_details.
• Streaming responses now surface reasoning_tokens under completion_tokens_details.
• Pricing is unchanged — these fields are reporting-only so callers can reconcile their own ledger 1
  against ours.

[2026-06-18]

Changed — context_management now requires the anthropic-beta header

• /v1/messages: sending context_management without the anthropic-beta header now returns a 400 at the gateway edge (previously the field passed through and silently degraded). Add the anthropic-beta header when using context_management.

Changed — Rejected sampling params stripped on native Claude paths

• Opus 4.7 / Opus 4.8 reject temperature, top_p, and top_k. These params are now stripped on the native /v1/messages path so the request stays native (keeps cache fields and the native msg_ id) instead of falling back to a lossy path. The params are dropped, not honored, for those two models.

Changed — Model identity injection defaults off

• The system-prompt model-identity directive is no longer appended by default (it had been adding ~40–130 phantom billed input tokens). Reconciled input_tokens now matches exactly what you sent.

[2026-06-17]

Models — GLM-5.2 added as Z.AI flagship

• Added zai/glm-5.2 — 1M-token context, 262,144 max output, pricing $1.40/M in · $4.40/M out ($0.26/M cached). Now featured on the homepage models strip.
• gpt-5.1 / gpt-5-pro routing corrected so those ids serve the genuine models rather than a substitute.

[2026-06-12]

Models — Kimi K2.7 added as Moonshot flagship

• Added moonshot/kimi-k2.7 — 256K context, image + video input, returns reasoning_content. Pricing $0.95/M in · $4.00/M out. Earlier Kimi versions remain routable but are hidden from the public list.

Added — Video input_type param + usage block on Seedance responses

• /api/v1/videos/generations accepts an optional validated input_type enum (text | image | first_last_frame | reference); when supplied it must match the seed fields actually sent, otherwise the request returns a 400 before any charge.
• Completed video poll responses now include a usage block (total_tokens).

Added — Per-call completion id for 1

reconciliation

• Chat responses now carry the completion id (chatcmpl-… / msg_…) and image/video responses carry a per-call id, surfaced publicly as id, so payers can join their ledger to ours.

[2026-05-24 — late]

Changed — Virtual Portrait enrollment dropped to $0.01 (parity with RealFace)

• POST /api/v1/portrait/enroll: $0.50 → $0.01 USDC. Storing an enrolment costs us nothing, so per-enrollment margin was always trivial; the old $0.50 was friction without offsetting cost. Now matches RealFace pricing — both are 1¢ to encourage adoption, and we make our money on the downstream Seedance per-clip revenue.
• All UI, docs, video-playground tooltips, and asset-storage comments updated to reflect the new price.

[2026-05-24]

Added — RealFace enrollment (real-person likeness, no KYC)

• POST /api/v1/realface/init + POST /api/v1/realface/enroll — full RealFace flow now live. The rights-holder scans a QR on their phone and completes a brief liveness check (nod + blink, ~1 minute), BlockRun's backend uploads the face photo and waits for an upstream biometric match, returns a ta_xxxxxx id usable as real_face_asset_id on Seedance 2.0 / 2.0-fast. $0.01 USDC per enrollment (promotional pricing), one-time. See RealFace API reference.
• No KYC required. No government ID, no account login, no name verification. Only a face-match between the live H5 capture and the photo you supply. Biometric data flows phone → upstream identity service directly, never touches BlockRun servers.
• GET /api/v1/realface/status?groupId=… — free polling endpoint, used by the studio UI to detect H5 completion.
• GET /api/v1/wallet/<address>/realfaces — symmetric to /portraits, lists a wallet's enrolled RealFaces.
• blockrun.ai/studio/realface — web UI: connect wallet → upload image URL → click Start → QR + countdown → poll for H5 completion → sign payment → copy ta_xxx.
• Video playground dropdown now combines Virtual Portrait + RealFace assets into a single real_face_asset_id selector, grouped by type.
• Header Models dropdown adds "RealFace Studio".

Reversed — earlier 2026-05-22 "RealFace removed" entry

• The 2026-05-22 changelog entry "Removed — RealFace" was based on an integration that was blocked by an upstream webhook bug at our inference partner. The partner shipped a fix on 2026-05-24; we verified the end-to-end flow (group active after H5 → photo upload → ta_xxx active in 6s → Seedance video generated in 86s using the enrolled face). The original conclusion that RealFace requires KYC was wrong — actual flow is liveness-only (nod + blink), no ID or account binding.

[2026-05-22]

Added — Virtual Portrait enrollment (no KYC)

• POST /api/v1/portrait/enroll — $0.50 USDC per enrollment, x402-gated. Registers an AI-generated character image as a Virtual Portrait and returns a ta_xxxxxx id usable as real_face_asset_id on Seedance 2.0 / 2.0-fast. Settlement happens only after enrollment succeeds — no charge on upload failure. See Virtual Portrait API reference.
• GET /api/v1/wallet/<address>/portraits — lists the portraits a wallet has enrolled. Free, rate-limited (20 req/hr/IP, shared with the wallet-reconciliation bucket).
• blockrun.ai/studio/portrait — web UI: connect wallet → name + image URL → sign → copy ta_xxx. Existing portraits listed below the form.
• Video playground at /models/bytedance-seedance-2.0-fast (and 2.0 Pro) now shows a dropdown of the connected wallet's enrolled portraits instead of asking users to paste a ta_xxx manually.
• Header Models dropdown now exposes "Portrait Studio".

Removed — RealFace (real-person likeness) is no longer offered (reversed on 2026-05-24, see latest entry)

• BlockRun does not offer real-person likeness in video generation. This conclusion was wrong — upstream had a webhook bug that made the group never transition to active; once fixed, the flow works without KYC. RealFace was reinstated on 2026-05-24.

Changed — Seedance defaults bumped to 720p + synced audio

• /api/v1/videos/generations now defaults resolution: "720p" for all Seedance models (was 480p). Override by passing an explicit resolution (360p / 480p / 720p / 1080p; 1080p is the max — 2K/4K are rejected with a 400).
• /api/v1/videos/generations now defaults generate_audio: true for text-to-video Seedance calls (false for image-conditioned or real_face_asset_id calls, where audio defaults are off per upstream guidance). Pass an explicit boolean to override.
• Output now matches the model maker's own output — same resolution, same multimodal audio path. Per-clip cost ≈ 2× the old 480p baseline because Seedance meters by tokens and 720p doubles the per-second count.
• Updated pricing table: $0.46 / $1.19 / $1.49 per 5s 720p clip for 1.5-pro / 2.0-fast / 2.0 (was $0.22 / $0.57 / $0.71 at 480p, text-to-video).
• tokensPerSecondAtDefault updated to 20,256 in the model catalog (was 10,128).

Fixed — Seedance copyright-filter UX

• The video playground now distinguishes Seedance's content-policy rejections from generic errors. On rejection, users see a clear "You were not charged" message + suggestions (rephrase with generic descriptors, switch to Grok Imagine Video, or enroll a Virtual Portrait for consistent characters).

[2026-05-18]

Added — Surf crypto data partnership

• /api/v1/surf/* — 83 crypto data endpoints in three flat price tiers ($0.001 / $0.005 / $0.02), covering CEX prices, on-chain SQL (sub-second ClickHouse, 80+ tables), Polymarket/Kalshi/dFlow prediction markets, wallet labels (smart-money, KOL, exchange), social mindshare, AI-curated news, and project/fund/token/wallet search. Payment settles 1
  directly to Surf's treasury — BlockRun takes no margin in exchange for being the discovery + x402 layer. See Surf API reference.
• blockrun_surf MCP tool added to the BlockRun MCP server.
• Surf added as a partner in Ecosystem.

Added — Phone & Voice partnership (Bland.ai + Twilio)

• /api/v1/voice/call — outbound AI conversation calls. $0.54 flat per call (up to 30 min, default 5 min). Bland.ai upstream.
• /api/v1/phone/numbers/buy — provision a wallet-owned US/CA phone number for $5 (30-day lease). Twilio upstream. Ownership recorded in Firestore against the calling wallet — only that wallet can use the number as caller ID or renew it.
• /api/v1/phone/numbers/renew — extend lease by 30 days for $5.
• /api/v1/phone/numbers/list — list a wallet's active numbers.
• GET /api/v1/voice/call/{id} — free polling endpoint for call status / transcript / recording.
• See Phone & Voice API reference.

Added — RealFace for Seedance video (deprecated 2026-05-22 — see Removed)

• bytedance/seedance-2.0-fast and bytedance/seedance-2.0 now accept a real_face_asset_id parameter (ta_xxxxxx format) for face-reference image-to-video. Asset IDs are managed in the upstream RealFace catalog.
• Superseded by Virtual Portrait enrollment (2026-05-22) — the API field name is preserved, but the only supported asset type is now a BlockRun-issued Virtual Portrait (POST /v1/portrait/enroll). Real-person likeness is not offered.

Added — Brand: Franklin → Franklin Agent

• Renamed the autonomous-agent product across the entire site, docs, and SDKs. Repo (BlockRunAI/Franklin) and CSS unchanged; surface copy now consistently says "Franklin Agent" to disambiguate from the personal name.

Changed — Seedance pricing migrated to per-M-token

• Old per-second pricing ($0.03–$0.30 / sec) is deprecated. New per-M-token pricing aligned with our upstream provider:
  • seedance-1.5-pro: $4.32/M tokens (flat, 480p default)
  • seedance-2.0-fast: $11.20/M (text-only) or $6.60/M (with image input)
  • seedance-2.0: $14/M (text-only) or $8.60/M (with image input)
• Default resolution is 480p (upstream default). Effective rate at 10,128 tok/sec.

Added — Rate-limit transparency

• 429 responses now carry Retry-After (RFC-7231 standard) + X-RateLimit-Source: <provider> headers, plus body fields code: "RATE_LIMITED", source, retry_after_seconds. Clients can fail over to a same-tier model on a different provider, or honor the retry interval. See Rate Limits — there is no platform-side QPS/RPM/TPM quota on paid inference; upstream provider limits apply.

[2026-04-25]

Models

• Added openai/gpt-5.5 — released 2026-04-23, first fully retrained base since GPT-4.5. 1M context, 128K output, native agent + computer use. Pricing: $5.00 / $30.00 per 1M tokens.
• Promoted openai/gpt-5.5 to FEATURED_MODEL_IDS (replaces openai/gpt-5.4 as the OpenAI flagship surfaced on the homepage and discover pages).

SDK Updates

• Python SDK (blockrun-llm) 0.17.0: PREMIUM_TIERS["MEDIUM"] primary now openai/gpt-5.5; gpt-5.4 demoted to first fallback. estimate_cost baseline rebased from $2.50/$15 to $5.00/$30 to match the new flagship. __version__ and VERSION reconciled (was 0.16.1 vs 0.15.0).
• TypeScript SDK (@blockrun/llm) 1.11.0: catalog refreshed with openai/gpt-5.5. VERSION and package.json reconciled (was 1.9.0 vs 1.10.1).
• Go SDK (blockrun-llm-go) 0.4.0: RoutingPremium.TierMedium primary now openai/gpt-5.5 (was gpt-5.4).

[2026-04-17]

Added — xAI Grok Imagine

• Video generation — new endpoint POST /api/v1/videos/generations with xai/grok-imagine-video ($0.050/sec, 8-second clips, text or image-to-video). The endpoint handles xAI's async job + polling internally so callers get the MP4 URL in a single request. See Video Generation.
• Grok Imagine image models — xai/grok-imagine-image ($0.02/image) and xai/grok-imagine-image-pro ($0.07/image) added to POST /api/v1/images/generations.
• Grok 4.20 chat models (hidden, API-routable) — xai/grok-4.20-reasoning, xai/grok-4.20-non-reasoning, xai/grok-4.20-multi-agent. 2M context, vision-capable, $2/M in · $6/M out.

Added — Permanent media URLs via GCS backup

• Every generated image and video is now mirrored to BlockRun's Google Cloud Storage bucket and returned through a blockrun-hosted proxy URL (/api/media/...). Upstream URLs from imgen.x.ai and vidgen.x.ai are temporary — the returned url is permanent.
• Image/video responses now include:
  • url — permanent proxy URL (falls back to upstream if the backup step fails)
  • source_url — original upstream URL (for debugging / comparison)
  • backed_up — boolean indicating whether GCS mirroring succeeded

API Changes

• GET /api/v1/models now includes video models with billing_mode: "per_second" and pricing: { per_second, default_duration_seconds, max_duration_seconds }.

[2026-03-24]

API Changes

• /api/v1/models now returns extended metadata for each model: name, description, context_window, max_output, and categories
• Model categories indicate capabilities: chat, reasoning, coding, vision

Models

• 33 visible models across 9 providers (Base), 31 on Solana
• Added GPT-5.4 Mini ($0.75/$4.50, 400K context)
• Added GPT-5.4 Nano, GPT-5.3, GPT-5.3 Codex
• Added Claude Opus 4.6 (1M context), Claude Sonnet 4.6
• Added Gemini 3.1 Pro, Gemini 3 Pro Preview, Gemini 3.1 Flash Lite
• Added NVIDIA Kimi K2.5 (free tier)
• Removed xAI/Grok, Qwen, and legacy GPT-4 family from visible lineup (hidden, still routable via API)

SDK Updates

• TypeScript SDK (@blockrun/llm) v1.4.3: Model type now includes categories, contextWindow, maxOutput, properly mapped from API
• MCP Server (@blockrun/mcp) v0.4.2: blockrun_models shows context window and categories, smart routing updated to current models

MCP Smart Routing

• fast: Gemini Flash, GPT-5 Mini
• balanced: GPT-5.4, Claude Sonnet 4.6
• powerful: GPT-5.4, Claude Opus 4.6
• cheap: NVIDIA free models, DeepSeek, Gemini Flash
• reasoning: o3, o1, DeepSeek Reasoner

[Unreleased]

Added

• Continue Integration — ClawRouter merged as a native provider in Continue (32K+ ⭐), the open-source AI code assistant. Users can now select ClawRouter directly in Continue's provider settings for cost-optimized model routing. (commit becda53)

Coming Soon

• Streaming support
• Additional AI providers

[1.0.0] - 2025-12

Added

• Initial public release
• x402 v2 protocol support
• OpenAI models (GPT-4o, GPT-4o Mini, o1, o1-mini)
• Anthropic models (Claude Opus 4, Sonnet 4, Haiku 4.5)
• Google models (Gemini 2.5 Pro, Gemini 2.5 Flash)
• xAI models (Grok 4)
• DeepSeek models (V3, R1)
• Meta models (Llama 3.3 70B)
• Qwen models (Qwen 2.5 72B)
• Mistral models (Mistral Large)
• Python SDK (blockrun-llm)
• TypeScript SDK (@blockrun/llm)
• Base network support (USDC)
• CDP Facilitator integration

Technical

• EIP-3009 TransferWithAuthorization for gasless payments
• EIP-712 typed data signing
• Automatic retry with payment on 402 response
• OpenAI-compatible Chat Completions API

SDK Versions

Python SDK (blockrun-llm)

[0.1.0] - 2025-12

• Initial release
• LLMClient with chat() and chat_completion() methods
• AsyncLLMClient for async usage
• x402 v2 protocol support
• Error handling with APIError and PaymentError

TypeScript SDK (@blockrun/llm)

[0.1.0] - 2025-12

• Initial release
• LLMClient class with full TypeScript types
• chat() and chatCompletion() methods
• listModels() for available models
• x402 v2 protocol support
• Error handling with APIError and PaymentError

Version Policy

BlockRun follows Semantic Versioning:

• MAJOR - Breaking API changes
• MINOR - New features, backward compatible
• PATCH - Bug fixes, backward compatible

Release Notes

For detailed release notes, see the GitHub releases page for each repository.