September 2026 began with the densest 48 hours of frontier releases since the August wave. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1 at an unchanged list price with three breaking API changes. On September 2 Google released Gemini 3.8 Flash at the same introductory price as 3.7 Flash, with the end of that price now printed, alongside a Fairwind-gated Cyber variant; and Meta released Muse Spark 1.3 with a contributor tier listed the same evening.
For current API prices, access limits, and workload selection, see our maintained frontier model guide.
This ledger is the September edition of the dated tracker we have kept since May and August. Each row is dated on the vendor’s own announcement, priced from the vendor or the OpenRouter listing where the vendor publishes no price, and linked to the post where we covered it. Releases after September 3 are added to this page rather than given new posts, unless one forces a decision on its own.
- 01Three frontier models in two days, none of them a price cut.Fable 5.1 keeps Fable 5’s $10 and $50; only cache reads fell. Gemini 3.8 Flash keeps 3.7’s $0.75 and $3.75 and now carries a January doubling. Muse Spark 1.3 keeps 1.2’s $1.25 and $4.25. The cheaper rows at the start of the month were all late-August carry-ins; GPT-6 Luna at $0.10 input joined them on September 22.
- 02Three of the 24 September releases are gated or conditional.Mythos 5.1 is available only through Anthropic’s verification programmes, Gemini 3.8 Flash Cyber only through Google’s new Fairwind Program, and TypeSafe’s Jev only through early access. The first two have no public price; Jev’s is an OpenRouter listing.
- 03Listing dates and launch dates disagree by up to six days this month.IBM’s Granite 4.2 8B carries an August 25 release date on its model card and an August 31 OpenRouter listing. The frontier releases listed within hours. The rule stands: the vendor’s date is the date.
- 04Seven dated events are on the calendar, five of them still ahead.Mercury 2.5’s 80% discount ended September 8 at 07:00 UTC and Claude Code’s weekly-limit change landed September 14. Still ahead: the end of Eleven v4’s launch discount on October 12, MiMo-V2.5’s deprecation on October 21, the end of Solar Mini 4’s launch discount on October 22, Anthropic’s Enterprise Frontier Safeguards “this fall”, Gemini 3.8 Flash’s intro-price expiry on December 31 and the Gemini 3.8 TTS doubling on January 1, 2027.
01 — The ledgerThe September ledger.
| Date | Vendor · model | What it is | Price | Context | Primary source | Replaces |
|---|---|---|---|---|---|---|
| September 1–28 · checked against the primary named in each row | ||||||
| Sep 1 | Anthropic · Claude Fable 5.1 and Claude Mythos 5.1 | Same model, two safeguard levels; Mythos 5.1 gated to Anthropic’s trusted-access programmes. Three breaking API changes: forced tool choice removed, thinking bound one way by model, history edits invalidate thinking (enforced for accounts created on or after Aug 31) | $10 / $50 per M, unchanged from Fable 5; cache reads cut to $0.25 (was $1.00); batch $5 / $25 | 1M in, 128K out | anthropic.com announcement; platform.claude.com migration guide. OpenRouter listed Sep 1, 18:03 UTC | Fable 5 (still available) |
| Sep 1 | Perplexity · Hybrid Compute on Mac, with PPLX Qwen 3.8 27B | Computer tasks start in the cloud; sensitive steps and private-file access run locally on Apple silicon (macOS 15+, 24 GB minimum, 32 GB recommended) behind an on-device PII classifier; local work uses no cloud credits | Included in Perplexity Pro, Max and Enterprise | Local model; not applicable | perplexity.ai/hub (returns 403 to fetchers); tier-one same-day coverage used, as stated in our post | Cloud-only Computer tasks |
| Sep 2 | Google · Gemini 3.8 Flash | Third Flash release in six weeks by Google’s count. HLE-Verified 54.9% (vendor-run); Artificial Analysis index 59, cost per task $0.58 vs $0.40 for 3.7 Flash at the same per-token price. Live in the API, AI Studio, Antigravity, Gemini app, AI Mode and Sheets | $0.75 / $3.75 per M to Dec 31, 2026; $1.50 / $7.50 from Jan 1, 2027 (Google’s footnote) | 1,048,576 (OpenRouter) | blog.google launch post. OpenRouter listed Sep 2, 15:14 UTC | 3.7 Flash “remains fully supported” |
| Sep 2 | Google · Gemini 3.8 Flash Cyber | Cyber variant with looser mitigations, gated through the new Fairwind Program for governments, critical-infrastructure operators and core technology platforms. CWE-Bench pass@1 47.2% and other figures are vendor-run | Not published | Not published | blog.google launch post; deepmind.google/fairwind-program | 3.5 Flash Cyber (July) |
| Sep 2 | Meta · Muse Spark 1.3 | Asks clarifying questions, asks for help when stuck, confirms before consequential actions. Meta: ~20% fewer tool calls, ~25% fewer tokens than 1.2 (vendor-run). Artificial Analysis: index 61 at xhigh, $0.55 per task, ~57% more input tokens per task. Max reasoning “coming shortly” | $1.25 / $4.25 per M, cache hit $0.15, unchanged | 1,048,576 (OpenRouter) | research.meta.ai launch post; artificialanalysis.ai. OpenRouter listed Sep 2, 19:45 UTC | Muse Spark 1.2 (August) |
| Sep 2 | Meta · Muse Spark 1.3 Contributor | Discounted tier on which Meta trains on your traffic; the same arrangement as the 1.2 contributor tier | $0.10 / $0.20 per M, cached input $0.002 | 1,048,576 (OpenRouter) | OpenRouter listing Sep 2, 20:38 UTC (listing date; Meta’s launch post does not name the tier) | Muse Spark 1.2 Contributor |
| Sep 15 | TypeSafe AI · Jev 1.13 | “System One” model that returns a typed decision with probabilities instead of text; early access via a waitlist. Vendor claims 70 to 500 ms per call; no independent benchmark yet | $0.042 per M input tokens, no output charge, on OpenRouter (typesafe/jev-1.13) | 32,000 (OpenRouter) | TypeSafe AI announcement, Sep 15. OpenRouter listed Sep 18, 2026 with a “decisions” output modality | Not stated; TypeSafe’s first System One model |
| Sep 17 | PrismML · Ternary Bonsai 2 27B | Ternary-weight compression of Qwen3.8 27B: 1.76 effective bits per weight, 5.9 GB, 98.2% of the base model’s aggregate benchmark score (vendor-run); 262K context, text and image input | Open weights, Apache 2.0; no hosted price | 262,144 (vendor) | prismml.com/news/bonsai-2-27b; whitepaper on GitHub | Ternary Bonsai 27B (July) |
| Sep 17 | OpenAI · Astra for Law | Configuration, not a model. GPT-6 Astra combined with a legal search index (U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs) and custom instructions for legal analysis and writing. Offered first to selected law firms through Trusted Access in ChatGPT and Codex; eligible firms get Zero Data Retention on the API. Vals AI Legal Research Bench 54.0% overall correctness vs 38.7% for GPT-6 Astra with web search alone, at the highest reasoning effort for both (OpenAI-run on a private validation set) | Not published | Not published; GPT-6 Astra underneath | openai.com/index/astra-for-law | Nothing; “coming soon to the API”, no date and no price announced |
| Sep 17 | inclusionAI · Ming-Image-0.1-Design and Ming-Image-0.1-Design-Layer | Two 6B open-weight image models under MIT. Design is a text-to-image model for UI, infographics and posters that can output RGBA images with transparent backgrounds; Layer decomposes a flattened design into a requested number of editable RGBA layers from an image and a layer plan. Both cards name one 80 GiB GPU as the validated configuration. Dated by the Hugging Face cards, created September 17 and last modified September 23 | Open weights; the two OpenRouter routes list zero catalog prices, which is not a statement that use is free | Not applicable (image output) | huggingface.co/inclusionAI/Ming-Image-0.1-Design and -Design-Layer, read Sep 25. OpenRouter listed Sep 22, 16:48 UTC (Design) and Sep 23, 22:17 UTC (Layer) | Not stated on the cards |
| Sep 18 | Alibaba · Qwen3.8-Omni-Flash | Native omnimodal model: text, image, audio and video input, text output, 1M-token context. Qwen reports an average gain of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations and per-hour audio input more than 98% cheaper than that predecessor (vendor-run, Qwen’s own estimation method). API only; no open weights announced | $0.15 / $0.47 per M (cache hit $0.016) on Alibaba Cloud Model Studio, international deployment; not on OpenRouter at the Sep 18 check | 1,000,000 (vendor) | qwen.ai/blog?id=qwen3.8-omni-flash, dated Sep 18 on Qwen’s research listing; Model Studio model page updated Sep 17 | Qwen3.5-Omni-Plus for text-output omnimodal use |
| Sep 21 | xAI · Grok 4.7 | A new, larger base model than Grok 4.6, trained with a longer reinforcement-learning run on a task mix weighted toward problems that take many hours, and trained to understand the Grok Bot harness natively. xAI publishes a seven-benchmark comparison against Grok 4.6 High, GPT-5.6 Sol Max and Fable 5.1 Max; every figure in it is vendor-run, and the table below reproduces it with an evaluator column. xAI also reports 62.4% on LatchBio’s biosafety benchmark and 3.3% risky-prompt pass-through on HackerBench v0.3, its own benchmark (both vendor-run). No parameter count is published | $2 / $6 per M, “priced starting at” in xAI’s words; unchanged from Grok 4.6. A fast variant is served at “twice the output speed at twice the price” — vendor-stated, with no separate rate table published and no separate public route confirmed. No cache-read rate published on the announcement or the docs model page | 500,000 (docs.x.ai model page, read Sep 22); the launch announcement publishes no context window | x.ai/news/grok-4-7, dated Sep 21; docs.x.ai/docs/models read Sep 22 for the context window | Grok 4.6 (August 12); Grok 4.6 remains on xAI’s model list |
| Sep 22 | Anthropic · Claude Opus 5.5 | First model of the Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work and, at default settings, costs 40% less to run than Opus 5 — a comparison of Opus 5.5 at its default effort (medium) against Opus 5 at its default (high), all vendor-run; FrontierCode v1.1 54.4% in Anthropic’s headline table, and 54.6% at default (medium) effort in its chart text. Adaptive thinking always on; four breaking changes versus Opus 5 on the model page (thinking cannot be disabled, forced tool use errors, thinking blocks bound to model and conversation, the older computer-use tool rejected on the Claude API and Google Cloud). Tested pre-release by Frontier Design and METR, per Anthropic | $4 / $20 per M, cache read $0.20 (0.05x input — a new ratio), cache write $5 (5 min) or $8 (1 hour); batch $2 / $10; full 1M context at standard rates. Fast mode $8 / $40 on the Claude API only, research preview, “up to 2.5x” output speed, not available with the Batch API. All from Anthropic’s pricing and model pages, read Sep 25 | 1M in, 128K out; 300K out on the Batch API with a beta header | anthropic.com/claude-opus-5-5, dated Sep 22; platform.claude.com model overview (“Released September 22, 2026”), pricing and fast-mode pages. OpenRouter listed Sep 22, 16:32 UTC at the same rates | Nothing retired: Opus 5 stays at $5 / $25. Retirement of Opus 5.5 not sooner than Sep 22, 2027 per its model page |
| Sep 22 | OpenAI · GPT-6 Sol and GPT-6 Luna | Two GPT-6 tiers below Astra, priced, in OpenAI’s words, 50% below their GPT-5.6 promotional counterparts (Luna’s output falls further, $1.20 to $0.50); OpenAI frames the cut as passing on caching and inference savings. Vendor-run claims: on AutomationBench 1.0.6, Sol at xhigh beats Claude Opus 5 at max at 9% of its cost per task; DeepSWE v1.1 68.8% for Sol at max; OSWorld 2.0 offline 60.5% for Sol at xhigh. Competitor scores are taken from public reports, not rerun. Live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and in the API as gpt-6-sol and gpt-6-luna; not yet in Chat | Sol $2 / $10 per M, cached input $0.20, cache write $2.50; Luna $0.10 / $0.50, cached $0.01, cache write $0.125. Above 272,000 input tokens: Sol $4 / $15, Luna $0.20 / $0.75. Batch and Flex are exactly half; Fast mode is double (OpenAI pricing page, read Sep 25) | 1,050,000 in, 128,000 out (OpenRouter); OpenAI’s long-context band starts at 272,000 input tokens | openai.com/index/introducing-gpt-6-sol-and-luna, dated Sep 22, 18:00 GMT in OpenAI’s news feed; developers.openai.com pricing page. OpenRouter listed Sep 22, 18:12–18:13 UTC at the same rates | GPT-5.6 Sol ($4 / $20 promotional, at least through Nov 21, 2026 per OpenAI’s pricing page) and GPT-5.6 Luna; both stay on the pricing page |
| Sep 22 | Xiaomi · MiMo-V2.6-Pro, MiMo-V2.6-Flash and Pro-UltraSpeed | Two native multimodal open-weight models (MIT) plus an UltraSpeed serving mode of Pro at “up to 20x inference speed”. Xiaomi reports 46 on the Artificial Analysis Intelligence Index for Pro, less than six days of RL costing around $850,000 (Flash) and $2.62 million (Pro), and DeepSWE v1.1 rising 48.8 → 65.7 (Flash) and 58.4 → 72.6 (Pro) — every figure Xiaomi’s own. Weights, the technical report, a distilled 9B model and the RL research resources are open-sourced | Unchanged from V2.5, per Xiaomi: Pro $0.435 / $0.87, Flash $0.14 / $0.28, Pro-UltraSpeed $4.35 / $8.70 per M (overseas rates; cache-hit input $0.0036 / $0.0028 / $0.036). Batch is half for Pro and Flash; UltraSpeed and the V2.5 models have no batch tier | 1,048,576 (OpenRouter) | mimo.mi.com release note and pricing page, read Sep 25 (the note prints no date; our post dates the release Sep 22). Hugging Face MiMo-V2.6-Pro-RL card created Sep 21, 15:39 UTC; OpenRouter listed Sep 21, 20:07 UTC at Xiaomi’s overseas rates | MiMo-V2.5-Pro and MiMo-V2.5, deprecated at 10:00 Beijing time on Oct 21, 2026 |
| Sep 22 | Upstage · Solar Mini 4 | Compact agentic model: 35B total / 3B active parameters, 512K context, up to 128K output, fluent Korean with English and Japanese; training data cut-off February 2026 (all per Upstage’s console page) | $0.10 / $0.40 per M, cached input $0.01, on Upstage’s console and pricing pages; a 50% launch discount of $0.05 / $0.20 (cached $0.005) runs from Sep 22, 05:00 UTC through Oct 22 (UTC) per the pricing page. The OpenRouter route’s $0.05 / $0.20 (cache read $0.005) is that discounted rate, not half of list | 512K (Upstage); 524,288 (OpenRouter) | console.upstage.ai model page, “Released on 2026-09-22”; upstage.ai pricing page. OpenRouter listed Sep 23, 10:45 UTC | Not stated; Solar Mini remains on Upstage’s pricing page |
| Sep 22 | Upstage · Solar Decide | Decision model, not a chat model: served as a System One endpoint on Solar Mini 4, it answers typed questions with a choice, a score or a yes/no and a probability, and writes no text. Beta. Upstage’s model page prints no benchmark | $0.10 per M input, output free, on Upstage’s console; the OpenRouter route shows $0.05, displayed as 50% off | 512K (Upstage); 524,288 (OpenRouter) | console.upstage.ai model page, “Released on 2026-09-22”, read Sep 28. OpenRouter listed Sep 28, 10:50 UTC, on its alpha Decisions API | Nothing |
| Sep 22 | Google · Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS | Text-to-speech models, generally available on the Gemini API on Sep 22 per the changelog and announced on blog.google on Sep 23. Ship with a Voices endpoint, voice design and voice replication; replication through AI Studio is not available in Illinois, Texas, the EEA, the UK, Switzerland or India (Google’s footnote). Google’s blog counts “2,000+ production-ready voices”; the Gemini API changelog says “150+ prebuilt and custom voices”. Flash-Lite is the stated replacement for gemini-3.1-flash-tts-preview. No function calling, Live API or thinking | Paid tier, per M tokens, text in / audio out: Flash TTS $0.50 / $9.00 and Flash-Lite TTS $0.50 / $6.00 through Dec 31, 2026, doubling to $1.00 / $18.00 and $1.00 / $12.00 on Jan 1, 2027. Batch and Flex are half; Priority is 1.8x. Audio bills at 25 tokens per second | 8,192 in, 16,384 out (Gemini API serving limit) | ai.google.dev changelog, Sep 22; blog.google, Sep 23; Gemini API pricing and model pages, updated Sep 24. OpenRouter listed both routes Sep 23, 21:51 UTC at the promotional rates | gemini-3.1-flash-tts-preview and the two 2.5 TTS previews name these as replacements on Google’s deprecations page, with no shutdown date |
| Sep 23 | Fireworks AI · Ember-1 | Kimi K3 post-trained by Fireworks Research to cut reasoning length: Fireworks claims K3’s quality with 40% fewer tokens, and 35–50% shorter reasoning at unchanged accuracy across seven benchmarks and two customers’ production traffic (all vendor-run). A research preview: two-week serverless access, to be made permanent if community demand warrants. The first model from Fireworks Research, not Fireworks’ first model | $3 / $15 per M, cached input $0.30, on Fireworks serverless — the same rate as Kimi K3 on Fireworks, so any saving comes from fewer tokens. OpenRouter route $3 / $15, cache read $0.30 | 1,048,576 (OpenRouter); Fireworks’ model page says 1,040K | fireworks.ai/blog/ember-1, published Sep 23; Fireworks model pages for Ember-1 and Kimi K3. OpenRouter listed Sep 24, 00:07 UTC | Nothing; served beside Kimi K3 |
| Sep 23 | AionLabs · Aion 3.5 and Aion 3.5 Mini | Roleplay and storytelling “multi-model systems” built on the GLM family, in AionLabs’ own description; 262,144 context, 32,768 max output. Niche, and logged because the vendor dates it | $3.00 / $6.00 per M (cache read $0.75) and $0.70 / $1.40 (cache read $0.18), per AionLabs’ models API; OpenRouter lists the same | 262,144 | api.aionlabs.ai/v1/models and aionlabs.ai/docs/models, release date 2026-09-23; no announcement post. OpenRouter listed Sep 23, 13:42 UTC | Not stated; Aion 3.0 (May 5) and 3.0 Mini (May 14) stay listed with no replacement set |
| Sep 23 | Black Forest Labs · FLUX 3 Action | Open-weight 7B “world action model” for robot control, derived from the FLUX 3 backbone and pretrained with a strong emphasis on video. BFL reports 42.92% on RoboLab-120 against 36.8% for Cosmos 3 Nano (vendor-run, BFL’s own table). Three checkpoints — base, so101 and droid — are public on Hugging Face under BFL’s FLUX Kommunity licence. Not an image model; logged because BFL dates it | Open weights; no API price on the post | Not applicable | bfl.ai/blog/flux-3-action, dated Sep 23 on BFL’s blog index (the page’s own metadata says Sep 22); Hugging Face checkpoints created Sep 22 | Not stated |
| Sep 24 | Respan · Span-01 and Span-01 Lite | Behaviour classifier on OpenRouter’s Decisions API: scores behaviours you define in plain language as present, absent or not observable in a conversation, in one forward pass and with no text output. Respan’s headline benchmark ranks Span-01 first of nine models at 0.843 F1, while its per-domain table puts GPT-6 Sol ahead; both are vendor-run | Span-01 $0.02 per M input, output free; Span-01 Lite free (Respan’s announcement and the OpenRouter routes) | Not stated | respan.ai Span-01 announcement, dated Sep 24, and concept docs, read Sep 28. OpenRouter listed Sep 26, 01:52 UTC | Nothing |
| Sep 24 | Google · Gemini 3.8 Live with Live Avatar | Feature, not a model: a real-time speaking avatar coupled to Gemini 3.8 Live’s dialogue, with lip-sync and expressions across 97 languages by Google’s account. Custom avatars only through enterprise allowlisting; audio and video carry SynthID. In Gemini Enterprise from Sep 24. Excluded from the release counts | Not published on the post | Not applicable | blog.google (linked from deepmind.google), Sep 24, 2026 | Nothing; Gemini 3.8 Live itself carries a Sep 15 release date on Google’s deprecations page |
| Sep 22–24 | Alibaba Cloud Model Studio · Qwen3.8-Max-Prime and GLM-5.3-Prime | Serving tiers, not models: Alibaba’s “Fast mode (Prime)” sells an existing model with TPS “increased to 1.5 to 2 times that of the standard API” (Fast mode page, updated Sep 22). No announcement post; dated by the page updates. GLM-5.3-Prime is sold by Alibaba, not Z.ai — Z.ai’s pricing page publishes no Prime tier — and appears only on Alibaba’s Chinese Prime page. Excluded from the release counts | qwen3.8-max-prime $3.301 / $9.902 per M against $1.65 / $4.951 for qwen3.8-max on Alibaba’s USD pricing page (Beijing region, updated Sep 24): 2x, with no batch or caching discount and no free quota. glm-5.3-prime ¥16 / ¥56, cache hit ¥4 (Beijing), against ¥8 / ¥28 for glm-5.3: 2x. OpenRouter routes $4 / $12 and $2.80 / $8.80, both with Alibaba as the sole provider | 1,000,000 (OpenRouter) | alibabacloud.com Model Studio pricing page (Sep 24) and Fast mode page (Sep 22); help.aliyun.com Prime page. OpenRouter listed Sep 23, 19:20 and 21:40 UTC | Nothing; the base SKUs stay listed |
| Sep 25 | Perceptron · Mk1.5 | A model built to control embodied agents: text, image, video and audio in; text, points, boxes, polygons, clips and object tracks out, with tool calling. Vendor-run claims: leads three of the four video object segmentation benchmarks Perceptron measured, and localises hands in first-person video 50% better than the strongest Gemini model it tested | $0.15 / $1.50 per M on the OpenRouter route; no vendor price read | 36,864 (OpenRouter) | perceptron.inc/blog/introducing-perceptron-mk1-5, read Sep 29; the post prints no date, and Perceptron’s blog index dates it September 25, 2026. OpenRouter listed Sep 25, 16:11 UTC | Not stated |
| Sep 28 | Anthropic · Claude Sonnet 5.5 | Second model of the Claude 5.5 family. Anthropic says it costs up to 30% less per task than Sonnet 5 because it needs fewer tokens, and generates output more than 30% faster; Terminal-Bench 4.0 70.6% at max effort against Sonnet 5’s 10.3% (43.0% at high, the API default), all vendor-run. First Sonnet with cyber fallbacks, to Sonnet 5, and with reasoning-extraction classifiers; its thinking blocks are bound to the account that produced them. Five request settings return a 400 error, including thinking disabled, replaced by between_tools | $2 / $10 per M, cache read $0.20, cache write $2.50 (5 min) or $4 (1 hour); batch $1 / $5. Every rate identical to Sonnet 5 (Anthropic pricing page, read Sep 28) | 1M in, 128K out; 300K out on the Batch API with a beta header | anthropic.com/claude-sonnet-5-5 and system card, Sep 28; platform.claude.com models overview and migration guide. OpenRouter listed Sep 28, 18:04 UTC at the same rates | Nothing retired: Sonnet 5 stays at $2 / $10. Retirement of Sonnet 5.5 not sooner than Sep 28, 2027; Haiku 5.5 due “in the coming weeks” |
| Sep 28 | ElevenLabs · Eleven v4 and Eleven v4 Turbo | Expressive text-to-speech model and a low-latency Turbo variant for voice agents, in 90+ languages, with instant voice clones from 10 seconds of audio and Professional Voice Clones. Ranked first on Artificial Analysis’ Provider Voice Arena at 1319 Elo (read Sep 28; Turbo not listed). Vendor-reported: preferred in about 75% of ElevenLabs’ own blind tests; Turbo at about 100 ms median inference and about 150 ms median time to first speech | $0.08 per 1,000 characters for v4 and $0.04 for Turbo, the same as v3 and v3 Conversational; a 72% launch discount, to $0.022 and $0.011, runs until Oct 12 (ElevenLabs API pricing page, read Sep 28). Priced per character, not per token | 10,000 characters per request (v4) | elevenlabs.io/blog/eleven-v4, published Sep 28, 12:00 UTC; elevenlabs.io/docs/overview/models. No OpenRouter route | Nothing retired; Eleven v3 and v3 Conversational stay listed |
Each September row has a post behind it. The Anthropic row’s breaking changes are in our Fable 5.1 cost and breaking-changes post; the Gemini row’s January price is the subject of its own post; the Muse row’s two conflicting token numbers are explained in the Muse Spark 1.3 post; and the Perplexity row is covered in our Hybrid Compute post, which states that Perplexity’s own hub page could not be fetched and that tier-one same-day coverage was used instead. Two of the three rows added on September 18 are covered in our TypeSafe Jev post and our comparison of SSD streaming and ternary weights, which covers Bonsai 2.
The rows added on September 25 have posts behind them too. Opus 5.5’s 40% claim is taken apart in our Claude Opus 5.5 launch post; the GPT-6 pair is priced and benchmarked in our GPT-6 Sol and Luna launch post; MiMo-V2.6 in our post on what Xiaomi published about its training cost; the TTS pair in our Gemini 3.8 Flash TTS post, which carries the January doubling; Ember-1 in our Ember-1 post; the Alibaba Prime tiers in our census of speed tiers; and the Ming-Image pair in our post on open-weight editable graphics.
The September 21 Grok 4.7 row is the first in this month’s ledger whose vendor announcement leads on a benchmark table rather than on a price or a capability, so the table is reproduced below with the column the announcement does not print: who ran each evaluation. What changed, and what did not, is taken apart in our Grok 4.7 benchmarks and price post.
| Domain | Benchmark | Grok 4.7 xHigh | Grok 4.6 High | GPT-5.6 Sol Max | Fable 5.1 Max | Who ran it |
|---|---|---|---|---|---|---|
| Software engineering | CursorBench 4.0 | 46.3% | 40.4% | 41.7% | 51.8% | xAI |
| Software engineering | DeepSWE v1.1 | 71.0%* | 65.2% | 72.7% | 70.0% | xAI |
| Electrical engineering | EEBench | 64.0% | 53.0% | 39.4% | 56.4% | xAI |
| Multi-hour office work | AA Briefcase v1.1 | 1,657 | 1,546 | 1,487 | 1,678 | xAI |
| Multi-hour terminal work | Terminal-Bench 4.0 | 38.0% | 20.3% | 37.3% | 57.9% | xAI |
| Legal work | Harvey Legal Agent Benchmark | 19.6% | 15.8% | 2.5% | 6.7% | xAI |
| Clinical reasoning | HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% | xAI |
These are OpenRouter catalog facts read on September 22, 2026, one day after the launch. They are a routing-provider observation, not xAI’s published terms, and they do not agree with them. The route x-ai/grok-4.7 is priced at $1.60 per million input tokens and $4.80 per million output tokens, against xAI’s own $2 and $6. It carries a 500,000-token context and a catalog maximum output of 450,000 tokens. Reasoning is mandatory and cannot be switched off, with four effort levels — low, medium, high and xHigh — and high as the default. Prompts above 200,000 tokens reprice at twice the rate, to $3.20 and $9.60, which is the number that matters for agentic runs rather than the headline. The canonical slug behind the route is x-ai/grok-4.7-20260916, a date five days before the announcement; a slug date is not a release date, and this ledger dates the row on xAI’s announcement.
Eleven rows were added on September 25, every one re-read on the vendor’s own page that day: nine releases dated September 17 to 23, one product feature dated September 24 and one serving-tier row that is not a release. Two things in the OpenRouter catalog, observed September 25, are worth stating as catalog observations rather than vendor facts. The Solar Mini 4 route at $0.05 / $0.20 is not a routing-provider discount: it equals the 50% launch rate Upstage itself publishes through October 22 (UTC), against a $0.10 / $0.40 list. And the z-ai/glm-5.3-prime route is served by Alibaba alone, at twice Alibaba’s own GLM-5.3 rate; Z.ai publishes no Prime tier, so the row is filed under Alibaba, not Z.ai. The GPT-6 Sol and Luna routes match OpenAI’s published rates exactly, and their :batch routes are exactly half.
Five releases were added, each dated on the vendor’s own page: Upstage’s Solar Decide (September 22), Respan’s Span-01 (September 24), Perceptron’s Mk1.5 (September 25) and, on September 28, Claude Sonnet 5.5 and ElevenLabs’ Eleven v4 pair. Two of them show why a listing is never the date. Solar Decide was released on Upstage’s console on September 22 but listed on OpenRouter only on September 28, and Span-01 was announced on September 24 and listed two days later. One catalog addition is not a September release at all: the bytedance-seed/seed-audio-1-0 route, listed September 25, belongs to a model ByteDance Seed’s own post dates July 20, 2026, so it is excluded from the counts. Sonnet 5.5 and Eleven v4 each have their own post: Sonnet 5.5 and Eleven v4.
02 — ContextLate-August carry-ins.
Six releases in the last week of August are the price floor September is being read against, and two of them still have live dates inside September. They are recorded here once, dated on the vendor where a vendor date exists, so that a reader comparing Gemini 3.8 Flash’s $0.75 against GLM-5.3-Flash’s $0.075 has both rows on one page.
| Date | Vendor · model | What it is | Price | Context | Primary source | Replaces |
|---|---|---|---|---|---|---|
| August 25–31 · carry-ins | ||||||
| Aug 25 | IBM · Granite 4.2 8B | Dense 8B open model with reasoning effort levels; release date per IBM’s Hugging Face model card | $0.10 / $0.15 per M on OpenRouter | 131,072 | huggingface.co/ibm-granite/granite-4.2-8b. OpenRouter listed Aug 31, 20:06 UTC, six days after the card’s release date | Not stated on the model card |
| Aug 26 | Z.ai · GLM-5.3-Flash | MIT-licensed open-weight Flash model, revealed as the stealth ox-alpha model | $0.075 / $0.25 per M on OpenRouter | 1,310,720 | z.ai/blog/glm-5.3-flash. OpenRouter listed Aug 26, 13:59 UTC | GLM-5.2-class Flash |
| Aug 26 | Alibaba · Qwen3.8-Flash | Flash-class release covered in our August 26–31 days | $0.15 / $0.47 per M on OpenRouter | 1,000,000 | OpenRouter listed Aug 26, 19:37 UTC; vendor coverage in the August ledger | Earlier Qwen Flash tier |
| Aug 28 | Tencent · Hy4 preview | Preview of Tencent’s next flagship; covered in the August ledger | $0.834 / $2.501 per M on OpenRouter | 1,048,576 | OpenRouter listed Aug 28, 06:09 UTC | Hy3 |
| Aug 28 | Z.ai · GLM-5.3 (753B) open weights | Weights released under a bespoke licence after a two-week hold Z.ai attributed to safety evaluation of faster-than-expected cyber capability | Open weights; hosted pricing varies | Per host | z.ai/blog/glm-5.3 | GLM-5.2 |
| Aug 31 | Inception · Mercury 2.5 Preview | Diffusion language model; 1,107 tokens per second claimed (vendor-run). Comparison set named by Inception: GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, Claude Haiku 4.5 | $0.04 / $0.15 per M at an 80% discount until Sep 8, 07:00 UTC; list $0.20 / $0.75 | 260,000 in, 65,536 out | Inception site; OpenRouter listed Aug 31, 20:57 UTC | Mercury 2 (February) |
03 — CalendarDated events still to come.
Eight events already have dates. Five are price changes, one is a model deprecation, one is a usage-limit change covered in our Claude Code September 14 post, and one is a safeguards product Anthropic has announced but not shipped.
Mercury 2.5 discount ends
Inception’s 80% launch discount on Mercury 2.5 Preview expires. Anyone who routed a fast-cheap tier to it in the first week should have the list price in the forecast from this date.
Claude Code weekly limit change
The weekly-limit reduction Anthropic scheduled for September 14 lands. Details and the decision it forces are in our dedicated post; this row exists so the date is on the calendar.
Eleven v4 launch discount ends
ElevenLabs’ API pricing page runs the 72% launch discount on both v4 models until October 12. Budget voice agents on the list rate; the discount lasts two weeks.
MiMo-V2.5 models deprecated
Xiaomi’s pricing page names the date and time. The V2.6 replacements carry the same prices, so the migration is a model-name change; the deprecation calendar carries the row.
Solar Mini 4 launch discount ends
Upstage’s pricing page runs the 50% launch rate from September 22 through October 22 (UTC). The OpenRouter route currently passes the discounted rate through; expect it to double with the vendor.
Anthropic Enterprise Frontier Safeguards
Anthropic’s launch post says eligible customers can use Fable 5.1 under zero data retention until EFS arrives, phased “starting this fall.” Not available today; do not plan on it for a Q3 deployment.
Gemini 3.8 Flash intro price expires
Google’s launch-post footnote. The first frontier intro price this month with a published end date (the Gemini 3.8 TTS intro rates end the same day); the price index carries both numbers.
Gemini 3.8 TTS prices double
Google’s pricing page prints both periods for the two text-to-speech models released September 22. Batch and Flex stay at half of whichever rate is in force.
Two items are real but undated and therefore not rows: Muse Spark 1.3’s max reasoning level, which Meta says is “coming shortly after we finish additional safety testing,” and the Muse Spark open-weights release, listed in Meta’s roadmap with no date and tracked on our withheld-weights ledger. They become rows when they get dates.
04 — AnalysisHow to read the month so far.
The pattern across the first six September rows (September 1–2) is that per-token prices held and the mechanisms around them moved. Anthropic changed the cache-read price and three API behaviours. Google left the price and published its expiry. Meta left the price and changed the model’s behaviour so that it reads more on a fixed task. A buyer comparing the month by list price sees nothing happening; a buyer comparing cost per task, which Artificial Analysis publishes for the Gemini and Muse rows above, sees Gemini 3.8 Flash up about 40% on its predecessor and Muse Spark 1.3 up from $0.40 to $0.55. The frontier model API price index carries the per-token rows; this ledger is where the dates live.
The second pattern is gating. Two of those six rows require an application, and both vendors’ gated models are the versions with cyber safeguards loosened. That is now the standing shape of a frontier launch, and the eligibility rules differ enough by lab that they need their own table rather than a column here. The third pattern is that the cheap end of the market is entirely open-weight or diffusion and entirely late-August: GLM-5.3-Flash, Qwen3.8-Flash, Granite 4.2 8B and Mercury 2.5 between them cover $0.04 to $0.15 per million input tokens. September has since added a frontier-lab entry at that price: OpenAI’s GPT-6 Luna at $0.10 per million input tokens on September 22.
05 — MethodMethodology.
A dated ledger, not a roundup: each row is a verified release with a vendor primary, and the page is refreshed in place rather than replaced by later instalments.
- What was collected
- Model releases with a vendor announcement dated September 1–30, 2026, plus late-August releases whose pricing or dated events carry into September. For each: date, vendor and model, what it is, price, context window, primary source, and what it replaces.
- Sources
- Vendor announcements (anthropic.com, platform.claude.com, blog.google, deepmind.google, research.meta.ai, z.ai, IBM’s Hugging Face model card, Inception, TypeSafe AI, PrismML); Artificial Analysis for independent index and cost-per-task figures; live pulls of the OpenRouter models API on September 3 and September 18, 2026, filtered on created time, for listing times, prices and context lengths where the vendor publishes none; a further pull on September 22, 2026 for the Grok 4.7 route, one on September 25, 2026 for the rows added that day, and one on September 28, 2026 (21:28 UTC) for the five rows added on September 29, whose vendor pages were Upstage’s console, Respan’s announcement, Perceptron’s blog, Anthropic’s announcement and documentation, and ElevenLabs’ blog, models and pricing pages. September 25 vendor pages: Anthropic’s announcement and its model, pricing and fast-mode documentation; OpenAI’s announcement, news feed and pricing page; Xiaomi’s release note and pricing page; Upstage’s console and pricing pages; Google’s Gemini API changelog, pricing, model and deprecations pages and the DeepMind and Google blogs; Fireworks’ blog and model pages; AionLabs’ models API; Black Forest Labs’ blog and Hugging Face collection; Alibaba Cloud Model Studio’s pricing and Fast mode pages; the inclusionAI and Xiaomi Hugging Face cards.
- As-of date
- First collected September 3, 2026; the rows dated September 15, 17 and 18 were added on September 18, 2026. The Astra for Law row and the September 21 Grok 4.7 row were both added on September 22, 2026. The rows dated September 17 (Ming-Image) and September 22–24 were added on September 25, 2026, each read on the vendor’s page that day, except Solar Decide (September 22) and Span-01 (September 24), which were added on September 29, 2026 with Perceptron Mk1.5, Claude Sonnet 5.5 and Eleven v4. The page is dated September 2 for the releases it opens with; this row is the only statement of the collection dates, and it moves with each refresh.
- Configurations
- A vendor configuration of an existing model — a product name wrapped around a shipped model with retrieval, instructions or access terms — gets a row when the vendor announces it under its own name, marked “configuration, not a model” in its first line. It is not a release and is excluded from every release count on this page.
- Dating rule
- The vendor’s announcement date is the row date. An OpenRouter listing time is corroboration only; it can lead or lag the real date by days or months, as the Granite row shows. Where no vendor date could be located, the row says so and gives the listing time.
- Exclusions
- Post-trains of open-weight models by third parties (they are listed on the host’s page, not here); models with no public listing or announcement; benchmark-only previews without a name.
- Refresh
- Refreshed in place as September releases land; new rows, not new posts.
06 — ConclusionSame prices, moved mechanisms.
Two days, three frontier models, zero list-price cuts. The changes are in cache rates, token usage, expiry dates and who is allowed in.
The September ledger opens with the vendors holding their prices and changing everything around them: what a cache read costs, how many tokens a task uses, when an introductory price ends, and which customers get the version with the safeguards off.
This page will carry every dated September release as it lands, on the vendor’s date, with its primary. If a release forces a decision on its own it gets a post and a link from its row; if it does not, the row is the coverage.
