System Card, November 2026

Six things that changed

• OpenAI / Test this first if you run a mid model

  GPT-6.1 Terra arrived at $1 in, $5 out

  OpenAI's new mid tier (system card, Oct 26) accepts reasoning effort "none", writes well and reads screens and documents. On the Artificial Analysis index it scores 42 at high effort for 26 cents a task: 97 percent cheaper per task than Claude Sonnet 5.5 at its best setting ($10.02), the mid model most founders reach for, for a score 22 percent lower (42 vs 54). With reasoning off it starts answering in under a second, and Artificial Analysis measures about 1.5 times Sol's output speed (OpenAI claims 3x). If your product runs on last generation's mid model, this is the one to test first.

  ShareCopyLinkedInEmailCopy link

• Google / Finish any Argon trial before Dec 1

  Gemini 4 Argon went GA, then Carbon took first place

  Argon (model card) went GA Oct 13 at $2 / $10, Sonnet 5.5's list price, and costs 23 percent less per task than Sonnet at its best setting ($7.68 vs $10.02) for a score 7 percent lower (50 vs 54); the price doubles Dec 1. Carbon (Oct 28, no card yet) scores 2 percent above Opus 5.5 (59 vs 58) at 23 percent less per task ($7.76 vs $10.04) and tops LMArena Text, but is gated to cyber defenders for now.

  ShareCopyLinkedInEmailCopy link

• Artificial Analysis / Never compare old and new scores

  Artificial Analysis reset its index to v5

  On Oct 27 the index swapped in private test sets. Haiku 5.5 fell from 43 to 36 (16 percent) with no model change; Kimi K3.1 Pro from 55 to 46. The gap from the best closed model to the best downloadable open model is now 27 points (59 vs 32), 46 percent of the leader's score, against 12 points on the old scale. Never compare a September score with a November one.

  ShareCopyLinkedInEmailCopy link

• Mistral · Reflection AI · Meta · DeepSeek / Cheaper open models worth a trial

  Mistral, Reflection and Meta shipped what they said

  Mistral Large 4 (model card) landed Oct 29 at 63 percent lower cost per task than its own October preview ($2.06 vs $5.54), the top open model outside China. Reflection Beam (model card) shipped Apache 2.0 weights Oct 16 using half the output tokens of K2 Horizon for one point less. Muse Spark Open (model card), 816B parameters, is the top US open model at 25, 58 percent below the index leader. DeepSeek V4.1 Flash 1017 (model card) scores 32 at $0.60 a task, 94 percent cheaper than Opus 5.5 for 55 percent of its score, and runs 48 trillion tokens a week on OpenRouter.

  ShareCopyLinkedInEmailCopy link

• xAI · Mistral · OpenAI / Move off OpenAI TTS by Jan 6, 2027

  Grok Voice 3.0 leads speech to speech; OpenAI retires its TTS line

  Grok Voice 3.0 (release post, no card) tops the Artificial Analysis speech-to-speech index at 86.9, up 7 percent on Voice 2.0, with support-call accuracy up 27 percent (71.6 vs 56.5 on tau-Voice) at the same $0.08 a minute, though it answers slightly slower (0.81 s to first audio vs 0.70 s). Voxtral 3 (release post, no card) prices transcription at $0.0035 a minute, 22 percent below OpenAI's GPT Transcribe. OpenAI retires tts-1 and tts-1-hd on Jan 6, 2027 with a realtime-only replacement, so anything built on them needs a new home.

  ShareCopyLinkedInEmailCopy link

• Anthropic / Try Sonnet before Opus for documents

  Claude shipped no model and held the work boards anyway

  Opus 5.5 (system card) leads GDPval-AA (documents, slides, spreadsheets) by 28 Elo over Sonnet 5.5 at twice the list price, and Sonnet 5.5 lands within 2 percent of it on the Vals Terminal-Bench run (64.1 vs 65.2 percent). The newest entry on Anthropic's system card index is still Haiku 5.5 (Oct 7). API version 2026-10-21 removes temperature; Opus 5.5 fast mode is GA at 2x price ($8 / $40) for up to 2.5x output speed (Anthropic's claim, not yet measured independently).

  ShareCopyLinkedInEmailCopy link

The cards we read this month

Every claim above traces to one of these 18 documents. Where a vendor published no card, the release post stands in and is marked.

Platform notes read alongside the cards: OpenAI deprecations · OpenAI GPT-6 effort levels · Anthropic API 2026-10-21 migration · Gemini API changelog · Gemini zero data retention.

The release map

Every notable release since October 2024, one lane per lab. Closed labs sit above the line in solid dots; open-weight models sit below in hollow rings. Larger, labeled dots are generational releases: a new model family or major version rather than a point update. Click any dot for what it is best at and its source.

The pace is picking up. Our map holds 24 notable releases dated October 2026 and 15 in September, against 4 to 8 a month through the first half of the year; the two months together carry 12 generational releases (a new model family or major version, such as GPT-6.1 Terra or Gemini 4 Argon, not a point update or a new snapshot of an existing model), more than the first six months of 2026 combined (11). Between January and June a generational release landed about every two weeks; in September and October, about every five days. Some of that is our own tracking getting denser as the series matured, so read the ratio as direction, not a precise multiple; benchlm.ai keeps an independent release count that shows the same slope. For a founder the practical effect is that the "best model for the job" table below has a shorter shelf life than it did a year ago, which is the argument for a monthly review and a harness rather than a one-time choice.

Open the release map on its own page

Which model for which job

A founder's routing table. "Best" is the model independent boards favor for that job today; "value" is the one that gets most of the way there for a fraction of the price. Prices are USD per million tokens, input / output, list price on Nov 1, 2026.

Chatbots and voice: when speed matters

For a chatbot, a support agent or a voice assistant, the wait before the first word matters as much as the answer. Here is what to reach for when response time is part of the product. Figures are from Artificial Analysis (text) and its speech-to-speech board (voice), read November 3, 2026.

Vendor speed claims, checked: OpenAI says GPT-6.1 Terra completes tasks more than three times faster than Sol; Artificial Analysis measures about 1.4 to 1.6 times Sol's output speed. Anthropic says Opus 5.5 fast mode gives up to 2.5 times the output speed at twice the price; no independent measurement exists yet. No video or image model released this month published generation times we could compare, so this table covers text and voice only.

The price map

List prices per million tokens on Nov 1, 2026. Click a column heading to sort. Promotional prices with an end date are marked. Cost per task on the Artificial Analysis index is the fairer comparison, since verbose models spend more tokens on the same work. Speed and first-token time come from the Artificial Analysis leaderboard, read November 3, 2026, at the lowest reasoning setting it measured (in brackets): that is the setting a live chat would use. For reasoning models, first-token time includes thinking. n/a means we have no measurement.

What to watch

• Kimi K3.1 Pro weights, promised by Nov 4. If they land under a usable license, the strongest open model of the year becomes self-hostable.
• Gemini 4 Carbon's API date, and Argon's price step on Dec 1.
• Qwen4, after a "coming soon" slide on Oct 29 that listed Max, Flash, Plus and 27B.
• Whether OpenAI's proposed end of Cursor's direct model access on Nov 12 holds; GPT-6 models are still absent from Cursor.
• Labs retuning for the private v5 test sets; expect open-weight scores to move.

What developers actually run

Leaderboards measure what a model can do. OpenRouter's rankings measure what people pay to run, across hundreds of apps routing through one API, updated weekly. Both matter; they disagree more than you would expect.

Sources: openrouter.ai/rankings · OpenRouter announcements · OpenRouter zero data retention · open vs closed token share tracker · closed-weight usage, 51-week view.

Where founders are moving

Two questions come up on every founder call: should I switch models, and should I go open. The data says switching is at a record high and mostly happens for price, and that open weights are winning volume while closed models keep the spend.

What we learned routing models inside Lev

We are the team that builds Lev, and this review exists because we needed it ourselves. Lev is an AI cofounder: it evaluates a startup the way an investor would, researches its market and competitors, drafts the pitch and the investor list, and works through the founder's plan with them in chat. Each of those is a separate model call with its own feature label in our telemetry, so on the first of every month, the same run that reads the new cards also reads Lev's own October spend, feature by feature, and proposes what to test. The numbers below are Lev's; the lessons travel.

• Measure spend by feature before you measure by model

  Lev's investor-style evaluation, the report that scores a company on team, market, product and traction, was 35 percent of our monthly bill on its own. Founders run it early and often, so it earns the spend, but no model swap matters as much as knowing which feature to swap it on. If your LLM calls carry a feature label into your analytics, this takes one query.

• Web search fees are their own line item

  Seventeen percent of our October spend was search fees inside the market and competitor research steps, plus the step that classifies investors for a founder's target list, which paid two cents of search per record. Cache search results and cap calls per request before you touch the model.

• Reasoning tokens are where the mid tier gets expensive

  Market research and competitive research, our two biggest research steps, spent 69 to 79 percent of their output tokens on reasoning. An effort setting is a cheaper experiment than a model change, and it ships without a new vendor agreement.

• Check the effort levels a model accepts before you plan around it

  We had planned to drop GPT-6.1 Sol into the founder chat and the writing canvas, the surfaces where a founder is waiting on every reply and we run with reasoning off. Sol does not accept effort "none". Terra does. One line in the model docs changed the plan.

• Do not switch without a harness

  Every swap above is a bounded experiment: one call site, config A against config B, 20 to 30 real but anonymized inputs, a judge with a rubric and both answer orders, a pass or fail threshold. Without it every change is a guess, and the cheap model always looks fine in a demo.

• Data terms gate everything

  Zero-retention and no-training terms in writing are the gate for any new provider. In October Google turned prompt storage on by default for new Gemini projects; ElevenLabs trains on non-Enterprise data by default. Read the data page before the model card.

What Lev runs today, and what is on the bench for November

Lev ran entirely on OpenAI's GPT-5.4 family in October; no other provider appears in our telemetry. The table is honest about pace: none of the swaps proposed in October's edition has shipped yet, because the thing they all need, a quality harness to judge A against B on real founder inputs, was only built in the last days of October. Its first jobs are effort changes, not model changes.

How we run this for Lev, and how you can

System Card is a scheduled Claude job that runs for Lev on the first of each month; this edition came out of the November 1 run. The agent reads the previous edition, searches each lab separately for what shipped, opens every new system card or model card directly rather than the press release, and checks each vendor claim against the independent boards (Artificial Analysis, Vals, LMArena, OpenRouter). Then it turns to Lev: it pulls the month's production telemetry from PostHog, where every one of Lev's model calls carries a feature label (evaluation, market research, competitive research, pitch, investor classification, chat, canvas), and asks one question per feature: are we leaving quality, speed or cost on the table. It writes three things: this page, an internal report for the Lev product team with a proposed experiment per feature, and a release ledger that next month's run starts from. A person on the team reads the report before anything changes in Lev.

The table above is this month's output for Lev. The three that matter most are not model swaps in the usual sense: try GPT-6.1 Terra, not Sol, on founder chat and the canvas, because Terra runs with reasoning off; lower the effort setting on market and competitive research before touching the model, because reasoning tokens were most of their cost; and cache and cap web search inside research and investor classification, because search fees were a sixth of the bill. Each becomes a ticket in Linear with a keep-or-switch threshold, and runs on Lev's A/B harness before it ships. Next month's edition will say which ran, and what happened.

The tools are ordinary: Claude running on a schedule (a scheduled task with a skill file that holds the procedure) for the reading and writing; PostHog's LLM analytics for cost, tokens, reasoning share and latency by feature and model (any tracing tool that stores a feature label per call works, for example Langfuse or Braintrust); a 20 to 30 item golden set of real, anonymized founder inputs per feature; a small A/B script with an LLM judge, a rubric, and both answer orders; and Linear, where every proposed swap becomes a bounded ticket.

To do the same on your product:

1. Tag every model call with a feature name and log tokens, cost and latency. If you do nothing else, do this; everything below depends on it.
2. Once a month, pull spend by feature and by model, plus the two hidden lines: web search fees and reasoning tokens as a share of output.
3. Read the cards for what shipped, or read this page. Note price, effort levels supported, context, and the weak spots section; that is where the surprises are.
4. Pick one feature and one challenger. The biggest feature by spend and the cheapest model that plausibly matches it is the default first experiment.
5. Build the golden set from real inputs, run A against B, and judge blind with a rubric. Decide the pass threshold before you look at the results.
6. Switch only on a passing run, and only after the provider's zero-retention and no-training terms are in writing. Then log the experiment, including the ones that failed, so next month starts from evidence.

Go deeper

The people and boards we check our own reading against. If you only follow three, take the first three.

Method and sources

We read each vendor's system or model card directly, then check it against independent boards: Artificial Analysis, Vals, LMArena (via dated mirrors this month, since arena.ai served stale pages), Terminal-Bench, GDPval-AA, and the speech boards. Every price and score above was looked up on Nov 1, 2026. Vendor-only numbers are marked as such.

Cards and docs: GPT-6.1 Terra · Gemini 4 Argon · Gemini 4 Carbon · Mistral Large 4 · Kimi K3.1 Pro · DeepSeek V4.1 Flash 1017 · Muse Spark Open · Grok Voice 3.0 · OpenAI deprecations. Boards: Artificial Analysis · Index v5 · Vals Terminal-Bench 4.0 · GDPval-AA · AA speech to text · AA text to speech · OpenRouter rankings · Judge reliability study.