AI Coding Leaderboard.

As of July 2026, Anthropic's Claude Opus 5 holds the top SWE-bench Verified score at 97.0%, measured independently by vals.ai on a neutral bash-only harness one day after the model launched. Treat that #1 as a tie rather than a win: GPT-5.6 Sol sits at 96.2%, Grok 4.6 at 95.6% and Claude Fable 5 at 95.0% on the same harness, and every one of those gaps is inside the combined margin of error. SpaceXAI's Grok 4.6 left the verifying queue on Aug 13, 2026, one day after launch, when vals.ai scored it 95.60% ±0.92 straight into 3rd; xAI published no SWE-bench Verified number of its own, so that rank rests entirely on the independent run, and at $0.78 per test it is the cheapest of the top five. One row now carries a contradiction big enough to flag rather than bury: vals.ai has measured DeepSeek V4 Pro's GA build (0813) at 96.40%, which is 15.8 points ABOVE DeepSeek's own 80.6% claim and would place it 2nd overall. Every other vendor on this board overstates; this one appears to understate badly. We have left the ranked score on the vendor figure per our standing disclose-don't-rebase rule and flagged the row for a human call. Sol still leads where it is hardest to fake, solving 98% of the 1-to-4-hour tasks against Opus 5's 90%. Opus 5 is also the cheapest of the three at $5 per 1M input, half of Fable 5. The surprise behind them is Moonshot's Kimi K3 (2.8T MoE), which vals.ai now scores at 93.4%, third overall and the cheapest route to a 90%-plus score at $3 per 1M input. Moonshot published its weights on Jul 27, 2026, making K3 the highest-scoring open-weight coder on this board. Read that with one caveat: the release is 4-bit MXFP4 only, 96 shards totalling about 1,454 GiB, with no full-precision base checkpoint, so it is open for inference rather than open for training, and the Kimi K3 License is permissive but not OSI-approved. Effectively tied with it is OpenAI's GPT-5.6 Luna at 93.0% (added Jul 21, 2026, and the 0.4-point gap to K3 is well inside the error bars), which is the cost story of the whole board: vals.ai measured $0.21 per test against $1.15 for Sol and $2.05 for Fable 5. DeepSeek V4 Flash jumped from the verifying queue straight to 6th on Aug 5, 2026, when vals.ai scored the 0731 build at 88.8% SWE-bench Verified on the same bash-only harness, edging out Claude Opus 4.8's 88.6%. Its vendor-reported 82.7% Terminal-Bench 2.1 claim still cannot be independently checked, since it depends on a DeepSeek-built harness the company has not released. Claude Opus 4.8 follows at 88.6%, then Grok 4.5 at 86.6% — the value pick among the per-token-priced 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 1.2 ties Grok 4.5 exactly at 86.6% on the same harness, added Aug 7, 2026 the day after Meta shipped it with the Muse Code terminal agent; Grok holds the higher rank only on the SWE-bench Pro tiebreak, and Muse Spark 1.2's weakness is long work (64% on the 1-to-4-hour tier against 90% for Opus 5). Z.ai's GLM 5.2 enters at #11 on 82.8%, also measured by vals.ai, and it is the strongest fully-open-source option here, MIT weights on Hugging Face since Jun 16 at $1.40 per 1M input, now that Kimi K3 above it ships under a bespoke non-OSI license. Kimi K3 scores higher and its weights finally shipped on Jul 27, 2026, though under Moonshot's own license rather than an OSI-approved one. Meta's Muse Spark 1.1 sits 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. Two of Google's newly launched models left the verifying queue on Jul 23, 2026 when vals.ai evaluated both: Gemini 3.6 Flash enters 16th at 79.6% and Gemini 3.5 Flash-Lite 18th at 75.0%, at $0.30 per 1M input and within 3.8 points of the Flash tier costing five times more. It held the cheapest-ranked-model slot until Jul 30, 2026, when OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 per 1M, putting a 93.0% model below it on price. Two more independently scored rows joined on Jul 27, 2026 after a routine sweep found vals.ai carrying them and us not: Cursor's Composer 2.5 at 79.6%, statistically tied with Gemini 3.6 Flash and usable only inside Cursor, and OpenAI's GPT-5.6 Terra at 75.2%, the weakest of the three GPT-5.6 tiers on this benchmark despite costing more than Luna. Alibaba's Qwen3.8-Max joined the ranked board on Aug 10, 2026 at 85.6%, an unusual row in that its only number is the independent one: Alibaba shipped the 2.4T-parameter model on Aug 3 with no benchmark table at all, so it waited unranked until vals.ai ran it. Read it as tied with Muse Spark 1.2 and Grok 4.5 above it rather than behind them, and note it is the slowest ranked model here by a wide margin, 2502s per test against Grok 4.5's 199s. Eight models still stay unranked. Z.ai's GLM-5.3 is the newest, released Aug 14, 2026 as a post-training-only upgrade on GLM 5.2's base model with no SWE-bench Verified figure of any provenance; its headline claims are a chart-topping CyberGym 84.5% for vulnerability discovery and Terminal-Bench 3.0 28.3, and Z.ai is withholding the weights about two weeks for safety hardening. Google's Gemini 3.7 Flash launched Aug 13, 2026 at half the price of 3.6 Flash ($0.75/$3.75 per 1M through year end) with no SWE-bench Verified figure of any provenance: its launch claims are DeepSWE v1.1 65.3% and FrontierCode 1.1 Main 43.6%, both vendor-run, and the 3.6 Flash precedent argues for waiting, since Google claimed a similar DeepSWE jump there and vals.ai then measured a statistical tie with 3.5 Flash. The other six are Meta's Muse Glimmer, Meituan's LongCat-2.0, Cohere North Mini Code, Kwaipilot's KAT-Coder-Pro V2.5, Poolside's Laguna S 2.1 and Alibaba's Qwen3.8-27B, the dense 27B open-weight model released Aug 14, 2026 whose vendor card claims 61.7% SWE-bench Pro and 73.0 Terminal-Bench 2.1 but publishes no SWE-bench Verified figure. Nobody has published a SWE-bench Verified score for any of them and vals.ai has not run them, and we would rather show a blank than an estimate. Muse Glimmer is the newest of those, a 30B dense model Meta opened under Apache 2.0 on Aug 10, 2026 that runs inside a 24GB consumer GPU; Meta benchmarked it only against Gemma4-31B and Qwen3.6-27B in its size class and published no SWE-bench figure, and an open-weight Muse Spark 1.2 is promised for the coming weeks. Every number here is checked against an independent evaluation or a primary source before we print it.