Research · Agentic tracker

What agents can reliably do, from dated results.

An ongoing stream. Each series follows one benchmark over time; every point links to the primary source that reported it.

Coding agents

Software-engineering benchmark

Resolving real software issues in code repositories.

Unit: % resolved · higher is better

Model	Date	Score	Source
Claude Fable 5SWE-bench Verified (500-task benchmark). Lab-reported (Anthropic) as relayed by DataNorth; no independent verification; primary Anthropic page not text-readable. Scaffold not stated.	2026-06-09	95.0	Source ↗
Claude Opus 4.8SWE-bench Verified (500-task benchmark). Lab-reported (Anthropic system card Table 8.1.A) as relayed by Vellum; Anthropic's own release page shows scores only as an image. Opus 4.7 reported 87.6 in same source. Secondary relay of a primary figure.	2026-05-28	88.6	Source ↗
Gemini 3.1 ProSWE-bench Verified (500-task benchmark). Lab-reported (Google DeepMind), High setting, agentic coding single attempt.	2026-02-19	80.6	Source ↗
Claude Opus 4.6SWE-bench Verified (500-task benchmark). Lab-reported by Anthropic but figure as printed in Google's Gemini 3.1 Pro model card comparison table (Max effort, single attempt); Anthropic's own page gives 81.42% only in a footnote with a prompt modification (avg of 25 trials). Date = Anthropic release.	2026-02-05	80.8	Source ↗
GPT-5.2 ThinkingSWE-bench Verified (500-task benchmark). Lab-reported (OpenAI). xhigh reasoning effort. Last OpenAI release to report Verified.	2025-12-11	80.0	Source ↗
Gemini 3 ProSWE-bench Verified (500-task benchmark). Lab-reported (Google DeepMind) single attempt, as quoted in the Gemini 3.1 Pro model card comparison table (Feb 2026). Release date taken from the date in a Nov 2025 search-result URL (simonwillison.net 2025/Nov/18); not verified on a Google page.	2025-11-18	76.2	Source ↗
Claude Sonnet 4.5SWE-bench Verified (500-task benchmark). Lab-reported (Anthropic). Avg of 10 trials, full 500, 200K thinking budget, bash+edit tools with minor prompt addition; 82.0% with high-compute parallel scaffold (not used here).	2025-09-29	77.2	Source ↗
GPT-5SWE-bench Verified (500-task benchmark). Lab-reported (OpenAI). High reasoning effort; scored on 477 of 500 tasks (23 excluded as not reliably runnable on OpenAI infra) so not strictly comparable to full-500 rows.	2025-08-07	74.9	Source ↗

Computer-use agents

Computer-use benchmark

Completing tasks by operating a real desktop or browser.

Unit: % completed · higher is better

Model	Date	Score	Source
Claude Fable 5OSWorld-Verified. Leaderboard entry (Yutori mirror) reported by Anthropic with 1M context variant; date is leaderboard entry date (model released 2026-06-09).	2026-08-01	86.0	Source ↗
Intelligence-Indeed Agent (agent system, model not stated)OSWorld-Verified. Top leaderboard entry (Yutori mirror); a third-party AGENT SYSTEM, not a single lab model; scaffold unknown. Reported by Intelligence Indeed, not independently reproduced.	2026-07-25	90.2	Source ↗
Gemini 3.5 FlashOSWorld-Verified. Lab-reported (Google) as relayed by ThePlanetTools (secondary); announcement date 2026-06-24 per that article.	2026-06-24	78.4	Source ↗
Claude Opus 4.8OSWorld-Verified. Lab-reported (Anthropic system card Table 8.1.A) as relayed by Vellum; Pass@1. Anthropic changed its run method for this eval at this release and restated Opus 4.7, so earlier Anthropic rows are not like-for-like.	2026-05-28	83.4	Source ↗
GPT-5.5OSWorld-Verified. Lab-reported (OpenAI). xhigh reasoning; research environment. Same page quotes Claude Opus 4.7 at 78.0 (Anthropic later restated Opus 4.7 to 82.3-82.8 after changing its run method, so cross-lab comparisons are not like-for-like).	2026-04-23	78.7	Source ↗
Claude Sonnet 4.6OSWorld-Verified. Leaderboard entry (Yutori mirror) reported by Anthropic; date is leaderboard entry date (model released earlier in Feb 2026, exact date not verified). Human baseline from original OSWorld paper is 72.36%, so this is roughly human level.	2026-03-08	72.1	Source ↗
GPT-5.4OSWorld-Verified. Lab-reported (OpenAI). xhigh reasoning; research environment; human baseline 72.4% quoted on page. GPT-5.3-Codex 74.0 on newly introduced full-resolution image parameter (previously 64.7).	2026-03-05	75.0	Source ↗
Claude Sonnet 4.5OSWorld-Verified. Lab-reported (Anthropic). 100 max steps, averaged over 4 runs. Leaderboard mirror lists 62.9 (100 steps) dated 2025-10-31.	2025-09-29	61.4	Source ↗

General agents

Task-horizon measure

The length of task, in human time, an agent completes at a stated success rate.

Unit: minutes · higher is better

Model	Date	Score	Source
GPT-5.6 SolMETR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 11.3h (CI 5-40h) when cheating attempts are scored as failures (METR standard). 71h if cheating attempts discarded; over 270h if scored as successes. METR says none is a robust measurement; highest cheating rate METR has recorded. No 80% horizon given.	2026-06-26	678	Source ↗
Claude Mythos Preview (early version)METR time horizon, 50% success, 228-task suite. Independent (METR). LOWER BOUND: 'at least 16 hours' (95% CI 8.5-55 h); METR says measurements above 16h are unreliable because only 5 of 228 tasks are that long. Tested March 2026, update published 2026-05-08. Relayed by Startup Fortune; METR primary page not retrieved. 80% horizon not reported.	2026-05-08	960	Source ↗
Claude Opus 4.6METR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 11h59m on all tasks (7h11m on METR private tasks only). Noise-adjusted (SIMEX) 80% horizon about 76-85 minutes. METR note dated 2026-03-20; date column is Anthropic release.	2026-02-05	719	Source ↗
Claude Opus 4.5METR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 95% CI 170-729. 80% horizon not published on source page.	2025-11-24	320	Source ↗
GPT-5METR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 95% CI 117-480. 80% horizon not published on source page.	2025-08-07	214	Source ↗
Claude Opus 4METR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 95% CI 58-170. 80% horizon not published on source page.	2025-05-22	101	Source ↗
OpenAI o3METR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 95% CI 74-201. Model release date is from general knowledge, not shown on the METR page (verify). 80% horizon not published.	2025-04-16	121	Source ↗
Claude 3.7 SonnetMETR time horizon, 50% success, Time Horizon 1.1 task suite (228 tasks; software, ML, cyber). Independent (METR). Minutes of skilled-human task length. 95% CI 32-106. 80% horizon not published on source page.	2025-02-24	60	Source ↗

Tool use and orchestration

Tool-use benchmark

Following policy and calling the right tools across a multi-step customer-service style task.

Unit: % success · higher is better

Model	Date	Score	Source
GPT-5.5tau2-bench Telecom domain. Lab-reported (OpenAI), xhigh, ORIGINAL prompts with no prompt tuning (stricter than GPT-5.2/5.4 rows that used a helper instruction), so not like-for-like.	2026-04-23	98.0	Source ↗
GPT-5.4tau2-bench Telecom domain. Lab-reported (OpenAI), xhigh. NOTE: the GPT-5.5 page (below) lists GPT-5.4 at 92.8 under original prompts with no tuning, so the same model scores differently under different prompting.	2026-03-05	98.9	Source ↗
Claude Opus 4.6tau2-bench Telecom domain. Anthropic model, figure printed in Google's Gemini 3.1 Pro card (Max effort). Lab-reported by a competitor, not Anthropic's own page (table is an image).	2026-02-05	99.3	Source ↗
GPT-5.2 Thinkingtau2-bench Telecom domain. Lab-reported (OpenAI). xhigh effort; short helpful instruction added to system prompt; airline subset excluded.	2025-12-11	98.7	Source ↗
Gemini 3 Protau2-bench Telecom domain. Lab-reported (Google DeepMind) as quoted in the Gemini 3.1 Pro model card (Feb 2026). Release date from search-result URL, not verified on a Google page.	2025-11-18	98.0	Source ↗
GPT-5tau2-bench Telecom domain. Lab-reported (OpenAI). High reasoning effort.	2025-08-07	96.7	Source ↗

Research agents

Web-research benchmark

Finding and combining information across the open web.

Unit: % correct · higher is better

Model	Date	Score	Source
GPT-5.6 SolBrowseComp (1,266 questions), % correct. Lab-reported (OpenAI), max reasoning, single agent. Sol Ultra (four agents in parallel) 92.2. Same page lists Claude Mythos 5 at 88.	2026-07-09	90.4	Source ↗
GPT-5.5BrowseComp (1,266 questions), % correct. Lab-reported (OpenAI), xhigh.	2026-04-23	84.4	Source ↗
GPT-5.4BrowseComp (1,266 questions), % correct. Lab-reported (OpenAI), xhigh; search blocklist excludes sites carrying answers (blocklist updated vs GPT-5.2). GPT-5.4 Pro 89.3.	2026-03-05	82.7	Source ↗
Gemini 3.1 ProBrowseComp (1,266 questions), % correct. Lab-reported (Google DeepMind), search + Python + browse.	2026-02-19	85.9	Source ↗
Claude Opus 4.6BrowseComp (1,266 questions), % correct. Anthropic model, number as printed in Google's Gemini 3.1 Pro card (search + Python + browse). Anthropic's own page reports 86.8 with a multi-agent harness (conditions: web search+fetch, context compaction at 50k tokens, up to 10M tokens).	2026-02-05	84.0	Source ↗
GPT-5.2 ThinkingBrowseComp (1,266 questions), % correct. Lab-reported (OpenAI), xhigh effort. GPT-5.2 Pro 77.9; GPT-5.1 Thinking 50.8.	2025-12-11	65.8	Source ↗
OpenAI Deep ResearchBrowseComp (1,266 questions), % correct. Lab-reported (OpenAI). Agent trained for persistent browsing; OpenAI flags it was trained on data that teaches BrowseComp-style tasks. Date = benchmark publication; product launched earlier.	2025-04-10	51.5	Source ↗
OpenAI o1 (no browsing)BrowseComp (1,266 questions), % correct. Lab-reported (OpenAI BrowseComp paper); no browsing tool. Baseline for the benchmark launch. GPT-4o 0.6, GPT-4o with browsing 1.9.	2025-04-10	9.9	Source ↗

Scores are only comparable within one benchmark, one version and one reported setting. We note the setting when a source states it, and never merge different versions into one line.