AI Coding Leaderboard — Best AI Models for Coding, Ranked (2026)

the verdict

As of July 2026, OpenAI's GPT-5.6 Sol holds the top SWE-bench Verified score at 96.2%, measured independently by vals.ai on a neutral bash-only harness. Treat that #1 as a tie rather than a win: Claude Fable 5 sits at 95.0% on the same harness, and the 1.2-point gap is inside the combined margin of error. Claude Opus 4.8 follows at 88.6%, then Grok 4.5 at 86.6% — the value pick of the leaders, solving a task for roughly $2.31 of input against $5.64 for Opus 4.8 and $10.53 for Fable 5, at a third of the latency. Meta's Muse Spark 1.1 debuts at 82.0% and is cheaper still per solved task (~$1.52). Below that, DeepSeek V4 Pro, Gemini 3.1 Pro, MiniMax M3, Qwen3.7 Max and Kimi K2.6 cluster between 80.2% and 80.6%, three of them open weights — but those are vendor-reported figures on the vendors' own scaffolds, and where an independent number exists it runs 4 to 11.6 points lower, so they are not strictly comparable to the leaders above. Four models stay unranked with no number at all, including Moonshot's Kimi K3 (2.8T, the largest open-weight model yet announced): nobody has published a SWE-bench Verified score for them, and we would rather show a blank than an estimate. Every number here is checked against an independent evaluation or a primary source before we print it.

Fig 1 · benchmark SWE-bench Verified — the share of real, human-validated GitHub issues an AI resolves end-to-end. Only models whose scores we have confirmed are charted; models still being verified are left off rather than estimated. Sources: vals.ai independent evaluations where available, otherwise maker reports — each row states which. Vendor-reported scores use the vendor's own scaffold and run high.

today's standings

#	Model	SWE-bench Verified	SWE-bench Pro	Input	Best for
1	GPT-5.6 Sol ↗OpenAI	96.2%	—	$5/1M	The highest independently measured coding score of any model, and it holds up on the long tasks: 98% on the 1-to-4-hour tier where most models fall apart.Independent (vals.ai, Jul 14 2026, mini-swe-agent bash-only harness): SWE-bench Verified 96.20% ±0.86 — the top score on the board. Verified Jul 17, 2026; it had been unranked since Jun 26 because OpenAI published no SWE-bench number of its own, and it still has not. Read the #1 with care: the 1.2-point lead over Claude Fable 5 (95.00% ±0.98) is inside the combined margin of error (~0.9 sigma, not significant), so the two are a statistical tie and we rank Sol first only because it scored higher. Where it does separate is task length — 98% on 1-4 hour tasks vs 93% for Fable 5. OpenAI's own Terminal-Bench 2.1 claim is 88.8% (Sol) / 91.9% (Sol Ultra). No SWE-bench Pro score published. Pricing $5/$30 per 1M.
2	Claude Fable 5 ↗Anthropic	95.0%	80.3%	$10/1M	Mythos-class flagship for long-horizon agentic runs: the model to reach for when a task spans hours and hundreds of tool calls and has to actually finish.Independent (vals.ai, Jul 14 2026, mini-swe-agent bash-only harness): SWE-bench Verified 95.00% ±0.98. Held the top score until GPT-5.6 Sol was evaluated at 96.20% ±0.86 on the same harness — a 1.2-point gap that is inside the combined margin of error (~0.9 sigma), so the two are a statistical tie and we rank Sol first only because it scored higher. SWE-bench Pro 80.3% uses Anthropic's own scaffolding and is contested. Restored Jul 1, 2026 after a 20-day export-control suspension. Pricing $10/$50 per 1M.
3	Claude Opus 4.8 ↗Anthropic	88.6%	69.2%	$5/1M	The hardest agentic refactors and long, autonomous multi-file tasks where every point of accuracy saves a human review cycle.Independent (vals.ai, Jul 14 2026, mini-swe-agent bash-only harness): SWE-bench Verified 88.6% ±1.42. Corrected Jul 17, 2026: we previously printed Anthropic's own "~86%" and claimed independent evals tracked it within ~1 point, which was wrong — the independent number is 2.6 points higher, and our methodology is to prefer the independent one. Run through the Claude Code harness instead of the bare bash agent, vals.ai measures 85.8%, a reminder that the harness moves these numbers as much as the model does. SWE-bench Pro 69.2% is Anthropic-reported.
4	Grok 4.5 ↗SpaceXAI (xAI)	86.6%	64.7%	$2/1M	The best value at the top of the board: it solves a SWE-bench Verified task for about $2.31 of input, less than half what the two models above it cost, and it is the fastest of the leaders.Independent (vals.ai, Jul 14 2026, mini-swe-agent bash-only harness): SWE-bench Verified 86.6% ±1.52. Verified Jul 17, 2026 — it launched Jul 8 with no Verified score and we had it unranked as "Opus-class, unproven"; the independent number now backs the Opus-class claim, landing it 2 points under Claude Opus 4.8 at 40% of the input price. On our price-per-solved-task metric that is ~$2.31 against $5.64 for Opus 4.8 and $10.53 for Fable 5, the cheapest of any model scoring above 85%. It is also quick: 199.6s mean latency per task versus 566.9s for Opus 4.8. SWE-bench Pro 64.7% and Terminal-Bench 2.1 83.3% remain SpaceXAI-reported. Priced $2/$6 per 1M. Still not available in the EU.
5	Claude Sonnet 5 ↗Anthropic	85.2%	63.2%	$2/1M	The best closed-model value — near-Opus scores at ~2.5× less, and the default daily driver for most developers.Vendor-reported (Anthropic), on Anthropic's own scaffold. Independent comparison: vals.ai's bash-only harness measures Sonnet 5 at 79.6% ±1.80, 5.6 points lower — a gap that reflects the harness as much as the model, so treat the 85.2% as a best-case number. Intro pricing $2/$10 per 1M through Aug 31, 2026, then $3/$15.
6	GPT-5.5 ↗OpenAI	82.6%	58.6%	$5/1M	OpenAI's strongest agentic coder, with the deepest tooling and ecosystem breadth of the closed labs.Verified score from vals.ai independent eval; Pro is OpenAI-reported (rivals flag possible memorization on Pro). Price: OpenAI list $5/$30 per 1M (cached input $0.50).
7	Muse Spark 1.1 ↗Meta	82.0%	—	$1.25/1M	Meta's first paid model, and a genuine value pick: a top-10 verified score for $1.52 per solved task, cheaper per result than every model ranked above it.Independent (vals.ai, Jul 14 2026, mini-swe-agent bash-only harness): SWE-bench Verified 82.0% ±1.72. Verified Jul 17, 2026; Meta published no SWE-bench number of its own and still has not. Do not confuse this with the earlier Muse Spark, which vals.ai scores 74.4% — 1.1 is a 7.6-point improvement over its predecessor. Price-per-solved-task ~$1.52, the lowest of any model in the top ten. Weakness is long tasks: 52% on the 1-4 hour tier against 74% for Claude Opus 4.8. Launched Jul 9 2026 by Meta Superintelligence Labs; $1.25/$4.25 per 1M with $20 starter credits.
8	DeepSeek V4 Pro open ↗DeepSeek	80.6%	55.4%	$0.435/1M	The cheapest frontier-class coder — top open-weights score at ~11× less than Opus. Best pick when cost or self-hosting rules.Vendor-reported, as aggregated by llm-stats (June 2026). Corrected Jul 17, 2026: we previously called llm-stats an "independent tracker", but llm-stats labels its own SWE-bench Verified table "Verified: 0 / Self-reported: 104" — every score on it is vendor-claimed, so this is a vendor number, not an independent one. No independent evaluator has run V4 Pro; vals.ai has evaluated the plain DeepSeek V4 (77.4%), which is a different model. Tied with Gemini 3.1 Pro on Verified, ahead on Pro.
9	Gemini 3.1 Pro ↗Google DeepMind	80.6%	54.2%	$2/1M	Google's strongest coding model today, with deep Workspace/Cloud integration. (A 3.5 Pro is expected but not shipped.)Vendor-reported (DeepMind) pass rate. No independent eval of this exact model; vals.ai has run Gemini 3.1 Pro Preview (02/26) at 78.8%, a preview build we do not treat as the same model. Ties DeepSeek V4 Pro on Verified, trails it on Pro. Price: Google list $2/$12 per 1M for context up to 200K (doubles above 200K).
10	MiniMax M3 open ↗MiniMax	80.5%	59.0%	$0.60/1M	Open weights with 1M context, multimodal input and computer use — beats GPT-5.5 on SWE-bench Pro at 5–10% of the cost.Vendor-reported at launch (Jun 1, 2026). Independent comparison: vals.ai's bash-only harness measures MiniMax-M3 at 75.0% ±1.94, 5.5 points lower, so treat the 80.5% as a best-case number.
11	Qwen3.7 Max ↗Alibaba	80.4%	60.6%	$1.25/1M	The best non-Claude score on the hardest benchmark — 60.6% SWE-bench Pro — built for long-horizon coding agents.Vendor-reported (May 20, 2026), and the widest vendor-versus-independent gap on this board: vals.ai's bash-only harness measures Qwen 3.7 Max at 68.8% ±2.07, 11.6 points below Alibaba's claim — on that harness it would rank far down the table rather than here. Read the 80.4% as a best-case, own-scaffold number and weight it accordingly. Proprietary — the open-weights sibling is Qwen3.6-35B. Price: Alibaba Cloud (Singapore) promo $1.25/$3.75 per 1M; list $2.50/$7.50.
12	Kimi K2.6 open ↗Moonshot AI	80.2%	58.6%	$0.95/1M	A top-three open coder whose 58.6% SWE-bench Pro beats several closed flagships.Vendor-reported (10-run average on Moonshot's own SWE-agent harness). Independent comparison: vals.ai's bash-only harness measures Kimi K2.6 at 76.2% ±1.91, 4 points lower. Price: official Moonshot API $0.95/$4 per 1M (cache-hit input $0.16).
13	Gemini 3.5 Flash ↗Google DeepMind	78.8%	—	$1.50/1M	Frontier-ish coding at Flash speed and price, with computer use built in as a native tool.vals.ai independent eval. See our decode of its native computer-use tool. Price: Google list $1.50/$9 per 1M (cached input $0.15).
verifying	Kimi K3 openMoonshot AI	—	—	$3/1M	The largest open-weight model announced to date (2.8T MoE, 16 of 896 experts active). Third on Artificial Analysis GDPval-AA v2 and second on the private AA-Briefcase agentic benchmark, so the independent signal is genuinely frontier-adjacent.Vendor-reported (Moonshot, Jul 16 2026): Terminal-Bench 2.1 88.3% on Moonshot's own KimiCode harness. No SWE-bench Verified or SWE-bench Pro score has been published by Moonshot or any independent evaluator, so unranked pending confirmation. Re-checked Jul 17, 2026: vals.ai's independent board covers Kimi K2.7 Code, K2.6, K2.5 and K2 Thinking but has not run K3, and llm-stats lists K3 with no Verified score. Expect a number once the weights actually ship. Independent placements that do exist: 3rd on Artificial Analysis GDPval-AA v2 (1,687) and 2nd on AA-Briefcase (1,527). Price: official Moonshot API $3/$15 per 1M (cache-hit input $0.30). Weights not yet downloadable.
verifying	LongCat-2.0 openMeituan	—	59.5%	$0.30/1M	A 1.6T Mixture-of-Experts agentic coder trained entirely on domestic Chinese chips; led OpenRouter usage in stealth as "Owl Alpha".Vendor-reported (Meituan, Jun 30 2026): SWE-bench Pro 59.5%, Terminal-Bench 70.8%. No SWE-bench Verified score or independent eval published yet, so unranked pending confirmation. Re-checked Jul 17, 2026: vals.ai has not evaluated it and does not cover Meituan at all, and llm-stats does not list the model.
verifying	Cohere North Mini Code openCohere	—	—	Free/1M	Private, self-hosted agentic coding for enterprises that cannot send source to a cloud API; runs on a single H100.Open-weight ~30B mixture-of-experts coder, free to use, runs on a single H100 with 256K context. Cohere published no SWE-bench Verified/Pro or independent coding eval, so it stays unranked pending confirmation (verifying). Re-checked Jul 17, 2026: vals.ai has not run this model. Worth noting for expectations — the only Cohere model on its board, Command A, scores 7.8%, by far the lowest of the 70 systems tested, though that is a general-purpose model rather than this code-tuned one.
verifying	KAT-Coder-Pro V2.5Kwaipilot (Kuaishou)	—	65.2%	$0.74/1M	Long-horizon agentic coding on a budget: a 72B-active MoE tuned for tool use, with a cheaper Air tier at $0.15/$0.60 on the same 256K context.Vendor-reported (KAT-Coder-V2.5 technical report, arXiv 2607.05471): SWE-bench Pro 65.2%, second only to Opus 4.8 at 69.2%, plus a best-in-test PinchBench 94.9% for agentic tool use, all run under a unified Claude Code harness. No SWE-bench Verified score published for V2.5, so it stays unranked pending independent confirmation. Re-checked Jul 17, 2026: vals.ai has not evaluated it and does not cover Kwaipilot/Kuaishou at all, and llm-stats does not list the model. Pricing confirmed at $0.74/$2.96 per 1M (Pro) and $0.15/$0.60 (Air).

We rank by SWE-bench Verified (500 real, human-validated GitHub issues resolved end-to-end), tiebroken by the harder SWE-bench Pro. A score is only printed once confirmed against an independent evaluation or the maker's primary source — and every row states which kind it is. Where both exist, we print both: one as the ranked score, the other in that row's note. We would rather show you the gap than ask you to trust our pick. Our independent reference is vals.ai, which runs every model itself through the same minimal bash-only harness (mini-swe-agent), so the models are compared on equal footing. That matters more than it sounds: SWE-bench scores a model and its scaffolding together, and vendors report using their own tuned scaffolds. Against vals.ai's neutral harness, the vendor claims on this board run 2.6 to 11.6 points optimistic. So rows marked vendor-reported are best-case numbers and are not strictly comparable to the independent ones — where we know the independent figure, we print it in the row's note. That is a deliberate choice and you should know we made it: on four rows (Claude Sonnet 5, MiniMax M3, Qwen3.7 Max, Kimi K2.6) an independent score for that exact model exists and is lower, and we still rank on the maker's published figure because it is the number that model is sold and quoted on. We disclose the independent one beside it instead of quietly restating the board on a single evaluator's harness choice. The honest consequence: positions that straddle the two regimes are approximate. Qwen3.7 Max is the sharpest case — it sits at #11 on Alibaba's 80.4%, and on the neutral harness its 68.8% would put it far down the table. Note that llm-stats, which we previously miscredited as an independent tracker, labels its own SWE-bench Verified table "Verified: 0 / Self-reported: 104"; it aggregates vendor claims. Models still being checked are marked “verifying” and shown without a number rather than estimated. Prices are per 1M input tokens on the standard API tier and can change — always confirm current pricing with the provider.

our picks

Best overallClaude Fable 5

Anthropic's Mythos-class flagship, topping SWE-bench Verified at a confirmed 95.0%, the highest of any model. The pick for the hardest coding and longest agentic tasks that must run unattended.

Best value (closed)Claude Sonnet 5

85.2% SWE-bench Verified at $2/1M — near-frontier coding at a rounding-error price. The default for most work.

Best budget / openDeepSeek V4 Pro

80.6% Verified at $0.435/1M with MIT open weights — frontier-class results at ~11× less than Opus.

Best for autonomous agentsClaude Fable 5

Built for long-horizon runs across hours and hundreds of tool calls, and the top-scoring model on SWE-bench Verified (95.0%) for delegated, repo-wide work.

Best free optionClaude Sonnet 5

It's the default model for free claude.ai users — frontier-class coding at no cost for everyday tasks.

Hardest-tasks dark horseQwen3.7 Max

Its 60.6% on SWE-bench Pro is the best non-Claude score on the benchmark that's hardest to game.

compare head-to-head

how the field got here

Fig 2 · history The frontier since we started tracking on Jul 2 2026: +10.2 points in 15 days, across 2 changes of leader. We plot one point per date the ranked data actually moved — not per edit — and the axis is to scale, so a flat stretch means nothing changed rather than that we stopped looking. Full series: JSON · CSV, CC BY 4.0.

2021GitHub Copilot preview Autocomplete-in-the-editor goes mainstream.
2023ChatGPT + GPT-4, then Cursor Chat-based coding and the first AI-native editor arrive.
Aug 2024SWE-bench Verified launches A human-validated benchmark of real GitHub issues sets an honest bar.
Oct 2024Claude 3.5 Sonnet hits ~49% Agents begin resolving real issues, not just snippets.
2025Terminal agents Claude Code and Codex CLI move AI out of the editor into the whole repo.
Apr–Jun 2026Open weights close the gap DeepSeek V4, Kimi K2.6 and peers cluster at ~80% Verified — for pennies.
2026Verified saturates in the mid-80s SWE-bench Pro and Terminal-Bench become the real differentiators.
Jul 2026Claude Fable 5 returns Restored after a 20-day export-control suspension; retakes SWE-bench Pro at 80.3%.
Jun 30 2026Meituan open-sources LongCat-2.0 A 1.6T MoE coder trained entirely on domestic Chinese chips; vendor-reported 59.5% SWE-bench Pro, awaiting independent eval.

Primary sources

BenchmarkSWE-bench — the real-GitHub-issue benchmark & leaderboard
Benchmarkvals.ai — SWE-bench Verified — independent third-party evaluations
BenchmarkSWE-bench Pro (Scale) — the harder long-horizon leaderboard
PaperSWE-bench (arXiv) — how the benchmark is constructed
PressGENZ TECH — Claude Fable 5 returns — restoration & SWE-bench Pro lead
MakerAnthropic — news — Claude Opus 4.8 / Sonnet 5 model cards & pricing
MakerOpenAI — GPT-5.5 — GPT-5.5 release post
MakerMoonshot AI — Kimi K2.6 — open weights & reported scores
MakerGoogle DeepMind — Gemini — Gemini model pages
PressVentureBeat — MiniMax-M3 debut — M3 launch scores & pricing
BenchmarkW&B ml-news — Qwen3.7-Max scores — Qwen3.7-Max benchmark table
PressGENZ TECH — Meituan LongCat-2.0 — 1.6T MoE coder on Chinese chips (vendor-reported)

cite & embed

Free to cite and embed with attribution. Raw data: JSON · CSV. Embedding this live table adds it to your site and credits GENZ TECH.

CiteGENZ TECH. (2026). AI Coding Leaderboard. https://genztech.blog/ai-coding-leaderboard/

Embed

<iframe src="https://genztech.blog/ai-coding-leaderboard/embed/" width="100%" height="470" loading="lazy" title="AI Coding Leaderboard by GENZ TECH" style="border:1px solid #26282b;border-radius:10px;max-width:560px"></iframe>

quick answers

What is the best AI model for coding right now?

As of July 2026, GPT-5.6 Sol has the highest independently measured score: 96.2% on SWE-bench Verified, run by vals.ai on a neutral bash-only harness. Claude Fable 5 is effectively level at 95.0% on the same harness — the 1.2-point gap is inside the margin of error, so calling either one "the best" overstates the evidence. Sol does pull ahead on long work, solving 98% of tasks that take a human 1 to 4 hours versus 93% for Fable 5. Claude Opus 4.8 follows at 88.6%, and Grok 4.5 at 86.6% is the value pick of the leaders at $2 per million input tokens.

What is the best open-source AI coding model?

DeepSeek V4 Pro leads open weights at 80.6% SWE-bench Verified for $0.435 per million tokens (MIT license), with MiniMax M3 at 80.5% (1M context, multimodal) and Kimi K2.6 at 80.2% just behind. Read that photo finish with care: all three are vendor-reported on the vendors' own scaffolds, and where an independent evaluator has rerun them on a neutral harness they land lower — MiniMax M3 at 75.0% and Kimi K2.6 at 76.2%. The ordering is probably right; the absolute numbers are best-case. Moonshot's Kimi K3 is the largest open-weight model yet announced but has no published Verified score, so we leave it unranked.

What is SWE-bench Verified?

SWE-bench Verified measures the share of 500 real, human-validated GitHub issues an AI resolves end-to-end. Scores climbed from ~49% in late 2024 to 96.2% in July 2026, so it is now close to saturated at the top and the harder SWE-bench Pro and Terminal-Bench separate the field better. One catch worth knowing: the benchmark scores the model and its agent scaffolding together, so a lab can raise its number by improving the harness rather than the model. That is why we rank on vals.ai's runs, which put every model through the same minimal bash-only harness, and why vendor-reported scores on this page sit 2.6 to 11.6 points above the independent ones.

What is the cheapest good AI coding model?

DeepSeek V4 Pro at $0.435 per million input tokens is the cheapest frontier-class coder (80.6% SWE-bench Verified, open weights), though that score is vendor-reported. If you want a top-tier score for the least money, the better question is cost per solved task: Grok 4.5 gets there for about $2.31 and Meta's Muse Spark 1.1 for about $1.52, against $5.64 for Claude Opus 4.8 and $10.53 for Claude Fable 5. Claude Sonnet 5 at $2 per million is still the convenient pick and is free as the claude.ai default.

Is Claude better than GPT for coding?

Not any more, and this answer changed on July 17, 2026. Claude held the top spot for months, but once vals.ai ran GPT-5.6 Sol on its neutral harness it posted 96.2%, edging Claude Fable 5's 95.0% — though by less than the margin of error, so the honest read is a tie at the top rather than an OpenAI win. Below them Claude still has the depth: Opus 4.8 at 88.6% beats GPT-5.5 at 82.6%, and the gap widens on SWE-bench Pro (69.2% vs 58.6%). Pick on cost and task shape, not brand: Sol is $5 per million against Fable 5's $10, while Grok 4.5 undercuts both at $2.

How often is this leaderboard updated?

We refresh it with every major coding-model release or newly confirmed score — it's checked on our twice-daily publishing runs. It was last updated on the date shown at the top of the page.

// maintained by GENZ TECH — latest tech, decoded