The GTM Model Leaderboard

Operator take · Same top five as last capture, but the scores were rescaled as Arena's vote volume grew. Claude Fable 5.1 Max still leads on Confirmed Success at 13.71%, with GPT-6 Astra Max second at 11.54%. Claude Opus 5 High and Opus 5 Max are now a near tie at 10.25% and 10.16%, so read #3 and #4 as one tier, and Fable 5 High holds fifth at 8.81%. Anthropic holds four of the five slots. The interval on the leader (±1.72%) now overlaps Astra's (±2.10%) less than before but the gap is only about two points, so this is a lead, not a lock. If your agentic GTM workflows run on Opus 5 today, Fable 5.1 Max is still the one worth a head-to-head test on cost and reliability, not just the top-line percentage.

Operator take · Google still holds two of the top five, but the two Gemini Flash entries traded places: 3.8 Flash High moved up to #3 (1496) and 3.7 Flash High slid to #4 (1495), a one-point gap that is well inside the margin of error, so treat it as a coin flip rather than a real change. The top two did not move: Fable 5 still leads Creative Writing, Text arena Overall, and Expert, which is the only model on this board clear enough to be worth a deliberate test. What is still worth noticing is which Gemini is winning here. Both Google entries in the top five are Flash models, the cheap fast tier, beating the Pro model at the writing job. If you have been paying Pro prices for draft copy on the assumption that bigger is better for prose, that assumption is not holding up. The workflow does not change: draft with a top-three model, then run your own voice pass. Judge the bake-off on edit distance from your own voice, because that is the only score that matters here.

Operator take · The most static board on this page, and it has now gone another three days without a single Elo point moving anywhere in the top five. GPT-5.6 Sol at xHigh still leads at 1257 with Claude Opus 4.6 Search four points back at 1253, which is a tie given intervals of 7 and 5, not a win. OpenAI holds first and third, Anthropic holds second, fourth, and fifth. The stillness is the signal: search-grounded quality is bounded by the index behind the model and by citation discipline, not by raw model IQ, so it does not lurch the way the reasoning boards do. First to fifth spans 27 points. That is not enough to justify moving your research workflow between stacks. Keep using whatever your tool already runs, Claude in Claude Code and Cowork, GPT in ChatGPT, and spend the effort on your source list instead.

Operator take · GPT-6 Astra still holds the top WebDev slot, 1793 against Claude Fable 5.1 Max's 1753, a 40-point gap that narrowed slightly from 42 last capture. Nothing is moving fast here. Opus 5 Max holds third at 1691, and GPT-6 Sol Max enters at fourth (1689), though its ±19 interval makes it a statistical tie with Opus 5 Max. Qwen 3.8 Max is fifth (1671) and Kimi K3 dropped out of the top five. If you build internal tools or Clay integrations on Claude today, there is no urgency to switch. Fable 5.1 Max is still excellent and stable. The Astra gap is worth a real evaluation only if raw WebDev score is what you are optimizing for.

Operator take · The Vision and Document boards held from last capture. On Vision, Fable 5 High still leads at 1310, with Qwen 3.8 Max (1302), Opus 4.7 High (1301), Opus 4.7 (1300) and Opus 4.6 High (1299) all within a few points, so the order below #1 is noise. On Document, Claude Opus 5 High leads at 1516, Claude Fable 5.1 Max is second at 1513, Opus 4.6 High and Opus 4.6 tie at 1507, and Fable 5 High is fifth at 1496. Practical read is unchanged: screenshots and charts go to Fable 5, and contracts and multi-page PDFs go to Opus 5 High.

Operator take · New #1. Claude Opus 5.5 now sits at the top of the Artificial Analysis Intelligence Index, and it takes the first three slots across effort levels: Max at 58, Xhigh at 56, High at 54. Claude Fable 5.1 follows at 53 for both Max and Xhigh. GPT-6 Astra, which tied Fable 5.1 at 53 last capture, no longer shows in the top five on the card. The card did not surface cost per task today, so last capture's pricing figures are not repeated here. The practical read: the effort setting matters as much as the model, since Opus 5.5 High (54) beats Fable 5.1 Max (53). For board-level calls where a wrong answer costs more than API spend, test Opus 5.5 at max effort against whatever you run now, and check the per-task cost before you commit.

Operator take · Ranked by fit, not a single benchmark. Celeris-1 reads 1,506 tokens per second today, inside the same noisy 1,200 to 1,900 band it has bounced around for weeks, and the runner-up slot now belongs to Mercury 2.5 at 770. Throughput readings on re-runs swing wider than the gap between most models here, so do not tune a pipeline to any single number. Latency and price leaders held: Gemini 2.5 Flash-Lite at 0.30s time to first token, Llama 3.1 8B at $0.02 per million tokens. Your Clay run is bounded by rate limits and row count long before tokens per second. The rule holds. At 10,000 rows a frontier model is a budget mistake: route volume work to a fast, cheap tier and reserve frontier models for the 5% of rows that matter.

Operator take · Raw window size is not recall quality. A 10M-token window that loses the middle of your call archive is worse than a 1M window that doesn't. These rows are held again rather than rewritten, for the fourth capture running, because the context-window card keeps returning different orderings below the top two and none of them has held long enough to publish. Llama 4 Scout at 10M and Grok 4.20 at 2M are the two constants across every read. What is independently confirmed today is the recall side: Opus 4.6 at high effort still ranks #1 on Arena's Longer Query board with Fable 5 at #2, and that is the number that actually predicts whether a long transcript survives the trip. Grok 4.6 keeps its slot at 500K, but note its 32-second time to first token: fine for an overnight transcript synthesis job, painful if you are waiting on it live. Test with a needle from your own transcripts before committing.