SWE-bench Verified Leaderboard — July 2026
Claude Opus 5 takes the top spot at 96–97% on SWE-bench Verified, edging out GPT-5.6 Sol and Claude Fable 5 as the frontier coding benchmark tightens at the top.
Results dashboard
Overall score
Results table
| Agent | Score | Time | Cost | Pass |
|---|---|---|---|---|
| Claude Opus 5 | 96 | — | — | Pass |
| GPT-5.6 Sol | 96 | — | — | Pass |
| Claude Fable 5 | 95 | — | — | Pass |
| Kimi K3 | 93 | — | — | Fail |
| GPT-5.6 Luna | 93 | — | — | Fail |
| Claude Opus 4.8 | 89 | — | — | Fail |
| Grok 4.5 | 87 | — | — | Fail |
The snapshot
The SWE-bench Verified leaderboard tightened dramatically in July 2026. Claude Opus 5, released July 24, took the top spot at 96.0% (Anthropic-reported) / 97.0% (Vals.ai independent run), edging out GPT-5.6 Sol at 96.2% (Vals.ai) and the previous leader Claude Fable 5 at 95.0%.
The remarkable story is price-performance: Opus 5 matches or beats Fable 5 at half the cost ($5/$25 vs. $10/$50), and GPT-5.6 Luna — OpenAI's cheapest tier at $1/$6 — scores 93%, higher than the previous Opus-tier model (4.8 at 89%). The price of near-frontier coding capability has collapsed.
What changed since june
| Model | June 2026 | July 2026 | Change |
|---|---|---|---|
| Claude Opus 5 | — | 96% | New entry, new leader |
| GPT-5.6 Sol | — | 96% | New (July 9 release) |
| Claude Fable 5 | 95% | 95% | Dethroned |
| Kimi K3 | — | 93% | New (July 16, weights public July 27) |
| GPT-5.6 Luna | — | 93% | New (cheapest tier, remarkable value) |
| Claude Opus 4.8 | 89% | 89% | Unchanged |
| Grok 4.5 | — | 87% | New (July 9) |
What this means for buyers
The frontier is now a cluster, not a single model. Five models score within 3 points of each other. Choosing on SWE-bench alone is no longer sufficient — look at SWE-bench Pro, Terminal-Bench, and your own workload.
Price-performance inverted. The cheapest tier (GPT-5.6 Luna at $1/$6) outscored last quarter's Opus-tier model. Routing by task difficulty is now the dominant cost strategy.
Open-weight caught up. Kimi K3 at 93% is within 3 points of the closed-weight frontier, with weights public. For self-hosting teams, the gap is functionally closed on this benchmark.
Beware benchmark gaming. A 2026 study showed SWE-bench Verified can be gamed to 100% via pytest hook manipulation. Read the methodology behind any score, not just the headline number. The scores above are from standard harnesses (Claude Code, Codex CLI, mini-SWE-agent), not adversarial setups.
Sources
- Anthropic Claude Opus 5 release (July 24, 2026)
- Vals.ai SWE-bench Verified leaderboard (updated July 22, 2026)
- Artificial Analysis Intelligence Index (July 2026)
- OpenAI GPT-5.6 release materials (July 9, 2026)
- Moonshot Kimi K3 release (July 16, 2026)