State of LLM Benchmarks (July 2026): 296 Evals Tracked

The July 14 registry contains 296 benchmark definitions across 10 categories. That count is the operator receipt behind this review; it replaces the 290 in the previous title rather than merely advancing the date.

Claude Mythos 5 leads overall, coding, and agentic work. Claude Fable 5 is second on all three. GPT-5.6 Sol is third on all three, though its overall and coding rows carry Estimated evidence. The repetition is the important finding in July: three independent ranking views now tell the same broad story.

The method changed too. BenchLM no longer treats an unreported benchmark as a zero or forces every model through an identical publication checklist. BenchAlign v5 estimates comparable capability, preserves uncertainty, and tells readers whether the result is Supported or Estimated.

Top models overall

Table 1 Download CSV

Rank	Model	Type	License	Score	Evidence
1	Claude Mythos 5	Reasoning	Proprietary	83.9	Supported
2	Claude Fable 5	Reasoning	Proprietary	83.7	Supported
3	GPT-5.6 Sol	Reasoning	Proprietary	82	Supported
4	Kimi K3	Reasoning	Pending	81	Supported
5	Claude Opus 4.8	Reasoning	Proprietary	78.3	Supported
6	Muse Spark 1.1	Reasoning	Proprietary	77.4	Supported
7	Grok 4.5	Reasoning	Proprietary	76.7	Supported
8	GPT-5.4	Reasoning	Proprietary	74.2	Supported
9	GPT-5.5	Reasoning	Proprietary	73.5	Estimated
10	Qwen3.7 Max	Reasoning	Proprietary	72.8	Supported

Mythos is the capability leader, but it is restricted. Fable is the leading generally available row. GPT-5.6 Sol is the closest non-Anthropic challenger, and its Estimated overall label matters: the position is useful, but less settled than a Supported row with comparable evidence breadth.

Claude Opus 4.8 remains fourth overall at 77.8. That corrects another misleading old story. Opus 4.6 is not Anthropic's frontier representative, and a comparison that leaves out Opus 4.8, Fable, or Mythos is historical by definition.

Coding leaders

Table 2 Download CSV

Rank	Model	Type	License	Score	Evidence
1	Claude Mythos 5	Reasoning	Proprietary	80.8	Supported
2	Claude Fable 5	Reasoning	Proprietary	80.6	Supported
3	GPT-5.6 Sol	Reasoning	Proprietary	78.9	Supported
4	Kimi K3	Reasoning	Pending	78	Supported
5	GPT-5.6 Luna	Reasoning	Proprietary	72.6	Supported
6	GPT-5.5	Reasoning	Proprietary	71.4	Supported
7	Claude Opus 4.8	Reasoning	Proprietary	70.6	Supported
8	Claude Opus 4.7	Non-Reasoning	Proprietary	68.4	Supported
9	Claude Sonnet 5	Reasoning	Proprietary	68.4	Supported
10	GPT-5.6 Terra	Reasoning	Proprietary	66	Supported

The coding order is not derived from one SWE-bench result. BenchAlign combines the coding evidence, normalizes each benchmark against its competitive field, and limits the influence of overlapping or repeated sources. That avoids equating 90 on an easier benchmark with 90 on a harder one.

The first four rows deserve different readings. Mythos and Fable have Supported evidence and lead. GPT-5.6 Sol is third with Estimated evidence. Opus 4.8 is fourth with Supported evidence. A team choosing between Sol and Opus should not read a 0.59-point score gap as certainty; it should run both on its own repositories.

Agentic leaders

Table 3 Download CSV

Rank	Model	Type	License	Score	Evidence
1	GPT-5.6 Sol	Reasoning	Proprietary	75.2	Supported
2	GPT-5.6 Terra	Reasoning	Proprietary	73.1	Supported
3	Muse Spark 1.1	Reasoning	Proprietary	68.6	Supported
4	Kimi K3	Reasoning	Pending	66.6	Supported
5	Claude Opus 4.7 (Adaptive)	Reasoning	Proprietary	65.9	Supported
6	GPT-5.5	Reasoning	Proprietary	65.8	Supported
7	Claude Mythos 5	Reasoning	Proprietary	65	Supported
8	Claude Opus 4.8	Reasoning	Proprietary	64	Supported
9	Claude Fable 5	Reasoning	Proprietary	63.9	Supported
10	GPT-5.3 Codex	Reasoning	Proprietary	61.2	Estimated

Agentic work is where the top three are best supported. Mythos, Fable, and Sol all carry Supported evidence. Their spread is only 1.6 points from first to third. The ranking is informative, but an operating decision should also measure intervention rate, tool-call recovery, latency, and the cost of a failed run.

Older rows such as o1 and o3 can still exist in the catalog without being sensible current recommendations. A benchmark database is also a historical record. The current rank should favor the current model generation when the evidence does.

Open-weight leaders

Table 4 Download CSV

Rank	Model	Type	License	Score	Evidence
1	MiniMax M3	Non-Reasoning	Open Weight	69.8	Supported
2	GLM-5.1	Reasoning	Open Weight	67.7	Supported
3	Inkling	Hybrid	Open Weight	67.5	Supported
4	GLM-5	Non-Reasoning	Open Weight	66.1	Supported
5	MiniMax M2.7	Non-Reasoning	Open Weight	64.1	Supported
6	GLM-5.2	Reasoning	Open Weight	64	Estimated
7	GLM-4.7	Reasoning	Open Weight	61.2	Supported
8	Gemma 4 31B	Reasoning	Open Weight	61.1	Supported
9	Qwen3.5-27B	Reasoning	Open Weight	60.7	Supported
10	DeepSeek V4 Pro	Non-Reasoning	Open Weight	60.7	Supported

MiniMax M3 leads the open-weight overall table at 69.8. GLM-5.1 and GLM-5 follow. The gap from Mythos is 14.05 points, so broad open-weight parity has not arrived. The operational comparison is narrower than that number suggests when privacy, fine-tuning, data residency, or high-volume serving makes an API model unsuitable.

The open-weight coding leader is different from the broad leader. GLM-5.2 and Kimi K2.7 Code sit at the top of the current open coding view. That is why the site keeps overall, coding, and agentic surfaces separate: the best general model is not automatically the best specialist model.

What BenchAlign v5 changes

The old ranking had two bad options for sparse models. Ignore missing benchmarks and a model could rise on a small set of favorable disclosures. Convert missing benchmarks to zeros and a model could be punished for what its lab chose not to publish. Requiring a fixed coverage threshold avoided some false precision, but it also left strong new models unranked.

BenchAlign v5 uses six controls:

1. Within-benchmark normalization. A score is judged against that benchmark's field, not against the raw percentage scale of another test.
2. Source and benchmark de-duplication. Closely related evidence cannot vote repeatedly as though it were independent.
3. Reliability weighting. Independent, reproducible, difficult-to-game sources carry more weight than weak or self-reported evidence.
4. Uncertainty-aware estimation. Missing observations widen uncertainty instead of turning into automatic failure.
5. Family consistency checks. Model-family information can stabilize sparse rows without forcing a newer or “Pro” label to win against contrary evidence.
6. Supported and Estimated status. Readers can distinguish a well-supported rank from a useful but less certain estimate.

No method makes the publication incentives disappear. It makes them visible and reduces the reward for selective disclosure.

Which benchmarks still matter?

Benchmarks matter when they separate current systems, resemble a consequential task, resist contamination, and publish enough detail to audit. Hard coding and agentic evaluations remain useful because model failures are observable and the field still has spread. HLE, GPQA, and MMLU-Pro remain more informative than saturated legacy knowledge tests. Grounded document and multimodal evaluations matter for users whose work is not plain text.

That does not mean every useful benchmark belongs in one universal score. Math and multilingual pages can remain lenses for readers and search engines while the core rank focuses on the six capability pillars. Multimodal evidence should inform versatility and relevant use cases without punishing a text-only configuration for not accepting images.

The July verdict

The current broad order is Mythos, Fable, GPT-5.6 Sol, and Opus 4.8. MiniMax M3 leads open weight. Mythos also leads coding and agentic work, with Fable second and Sol third.

The more important change is that a model can now rank without publishing the exact same benchmark set as every competitor. Its uncertainty remains visible. That is a better answer than either excluding it or pretending its missing rows are failures.

→ Live leaderboard · Coding ranking · Agentic ranking · Open-weight guide