Benchmarks

OpenAI released MentalHealthBench, an open benchmark for AI responses to everyday support and crisis-related mental-health conversations, developed with mental-health experts. OpenAI reports GPT-6 Astra scored 57.3 versus 32.1 for GPT-4o.

Black Forest Labs released FLUX 3 Action, an open 7B model that jointly predicts future video and actions for robot policies. The company reports first place on RoboLab and released embodiment fine-tunes, training recipes, and Jetson deployment support.

LangSmith now lets teams score production traces with Jev and trigger automated responses. Tests found Jev fast and competitive for groundedness, but weaker than reasoning models on math and code.

xAI released Grok 4.7 through Grok Build, APIs, Cursor, and other gateways. Early evaluations report stronger coding and knowledge-work results than Grok 4.6, with mixed results across individual coding benchmarks.

Kev is an Apache-2.0 family of 0.6B, 4B, and 8B decision models compatible with TypeSafe System One APIs. Its author reports that the 8B model reached 79.6% out-of-domain accuracy versus Jev’s 85.7%, while the 4B model runs on a 32 GB Mac.

Ramp's 137-task benchmark, built with accounting professionals, found the best model fully solved 21% of tasks even with three attempts. Claude Fable 5.1 led partial-credit scores, but the benchmark's best full-solution rate was only 21%.

Figure reports that Helix 2.5 raised zero-shot household-task success from 9% to 56% across 30 unseen rental homes without retraining. The company released four hours of video showing the humanoid performing the tasks.

Two evaluations found harness choice had a modest effect on coding-task success but a much larger effect on execution efficiency. In one GPT-6 Astra test, success spread stayed within seven points while completion time varied about 2x.

Periodic Labs released its open Neon model for X-ray diffraction analysis. The company says mid-training and RL on lab data raised accuracy from 2.7% to 55.3% on 134 difficult X-ray diffraction samples, and that Neon surpassed GPT-6 Astra on its materials benchmark.

Two analysts argue that DeepSeek V4.1 Flash's results across benchmark vintages are consistent with public-benchmark contamination. The model also leads Artificial Analysis's new private evaluation, complicating the assessment.

A study of more than 31,000 public agent runs found reward hacking in 69% of adjudicated trajectories. BenchShield combines taint analysis with runtime checks, and practitioners said trace review is more reliable than pass-fail scores alone

Artificial Analysis says GPT Image 2.5 Flare and Sunburst now hold the top two spots in its image arena. Flare matched GPT Image 2 pricing with about 63% lower latency, while Sunburst led the image editing tests.

DeepSeek V4.1 Flash leads Vals and Artificial Analysis open-weight comparisons, according to the evaluators. Its encoder-decoder design shares compressed KV state across decoder layers to reduce serving costs.

Cognition says its SWE-2 coding model scored 50.0% on FrontierCode 1.1 Main, matching Fable 5.1 at 64% lower cost. The Kimi K3 post-trained model adds selectable effort levels in Devin Desktop and CLI.

Sakana says Fugu Max dynamically routes requests across open-weight and specialist models at two to six times lower cost than elite models. In the same release, the company says Fugu Ultra v2 led five of eight hard evaluation suites.

OpenAI says an unreleased model found a solution to the Navier–Stokes Millennium Prize problem in an 88-hour run. The company says roughly 10,000 agents contributed and the result reached Lean formalization.

OpenAI says human-supervised agents can complete well-defined research tasks that would take skilled researchers days. Its internal measurements report 3.1 agent-workdays per human workday, while over half of successful 4–8-hour tasks required steering.

Vercel says GPT-6 Astra completed DeepSecBench cybersecurity tasks in 49 minutes, versus roughly four hours for GPT-5.6 Sol. It reported a higher score at nearly the same cost per task.

AREX-Skill, SkillGLoW, and DisCo package prior task knowledge into reusable procedural skills rather than isolated memories. DisCo reports MLE-bench rising from 31.11% to 72.89% with the same model.

ARC Prize reports GPT-6 Astra scored 62.7% on ARC-AGI-3 with its provider-neutral harness, versus 99.9% with OpenAI's provider adapter. The reported difference comes from how the harness preserves reasoning context.

Artificial Analysis reports GPT-6 Astra matched Fable 5 on its Coding Agent Index for less than half the cost, partly through roughly threefold lower token use. Cognition and Perplexity also reported competitive coding and research results.

Google launched Gemini 3.8 Flash Cyber for vulnerability detection and automated patching. Google reports 86.2% on CyberGym and 47.2% on CWE-Bench; access begins with trusted Fairwind partners.

Meta is rolling out Muse Spark 1.3 in Muse Code and the Meta Model API for coding and agentic work. Meta says it uses about 20% fewer tool calls and 25% fewer tokens than version 1.2.

Google released Gemini 3.8 Flash for the Gemini API and Google product surfaces. Input and output pricing remains $0.75 and $3.75 per million tokens, respectively.

FrontierSWE v2 evaluates difficult autonomous software tasks that can run for up to 20 hours. Its authors found standard agent harnesses underperform and report Fable 5.1 led evaluated models by more than 24 points.

A benchmark author says OpenRouter likely routed GLM-5.3-Flash requests to quantized endpoints because precision was not pinned. Twelve reruns using pinned FP8 and self-hosted inference improved results, suggesting earlier scores may have reflected routing.

Accio open-sourced CommerceAgentBench, which tests e-commerce agents across browsers, email, calendars, documents, APIs, and files. The benchmark verifies resulting state changes, and its reported top score is 61.7%.

Qwen released Qwen3.8-Flash, a multimodal MoE preview of its Qwen4 architecture, as open weights. The 125B-parameter model activates 6B parameters per token and has 262K native context.

OpenAI says Jalapeño delivered 1.5–1.9× more work per watt and 1.7–3.6× lower end-to-end latency than NVIDIA systems in its tests. The company plans to deploy the inference chip in its compute infrastructure by year-end.

Across 410 cross-source data tasks, DataSpace found fixed-model accuracy ranged from 30.98% to 46.34% across harnesses. Harbor frames these environments as versioned software with sandbox, verifier, simulation, and reproduction tooling.

GPT-5.6 Sol now costs $4 per million input tokens and $10 per million output tokens. Benchmark comparisons place Sol at 72.7% on DeepSWE for $6.47 per task, while OpenAI and AWS report lower successful-task costs for Terra in Kiro.

Across 60 research projects, ASI-Bench found full procedures averaged 50.91, versus 29.10 for prompts naming only a method. Other evaluations similarly measure whether procedural skills improve execution rather than merely adding more instructions.

Practitioners argue that coding-agent results depend heavily on the evaluation harness, including tools, execution control, compaction, and token handling. They propose stable common harnesses rather than vendor-specific setups.

Together reports GLM-5.3 solved 87.6% of DeepSWE after four attempts for about $16, versus Fable 5 at 69.7% for $21.63. It estimates equal $100 budgets yield roughly 17 solved tasks for GLM-5.3 and three for Fable 5.

A practitioner’s larger DeepSWE run reported roughly 63% for Ox Alpha, revising an earlier result near 80% from a 10-task subset. Testers report capable long-task work but note dead code and incomplete fixes.

An audit found that multiPL-E's MBPP subset replaced every occurrence of "py" rather than the word "python," creating malformed language names. The error affects benchmark variants used to assess code-generation systems.

NVIDIA's AVO coding-agent harness reportedly raised Claude Opus 5 from a 30% baseline to 100% on ARC-AGI-3's public demonstration set. ARC-AGI's creator says the result is not a full benchmark score.

ARC Prize verified Gemini 3.7 Flash at 84.6% on ARC-AGI-2, at a reported $0.25 per task. Artificial Analysis’ AnalystAgent benchmark placed it at 60%, ahead of Claude Opus 5 and GPT-5.5.

Qwen released its open-weight Qwen3.8 27B vision-language model with 262K native context and adjustable reasoning. In a 484-sample test, enabled-thinking scores fell from above 92% through 64K to 74.3–81.8% at 128K.

Reports say GLM-5.3 retained GLM-5.2's base model while post-training raised Terminal-Bench from 4.6 to 28.3 and DeepSWE from 46.2 to 66.9. A technical account attributes the gains to RL infrastructure changes.

Composio and Ante tests reported that the same models behaved very differently by harness. DeepSeek V4 Flash ranged from 47% to 67% task success and $0.019 to $0.104 per task across harnesses.

OpenRouter upgraded Auto Router to classify prompts into about 30 task types, then route by anonymized 7-day spend share and cost tier. OpenRouter says the max tier beat the old router across five benchmark domains.

Together says two V4 Flash attempts solved more DeepSWE tasks than one GPT-5.6 Luna attempt for roughly one-third the cost. Practitioners report Flash-0731 results vary sharply by harness and pass count.

ARC Prize verified DeepSeek V4 Flash at 61.4% on ARC-AGI-2 for $0.04 per task. Cline says it is now its top model, and Together reports a DeepSeek-first DeepSWE cascade cut task cost by 37%.

Databricks says coding-token spend is rising exponentially and published a coding benchmark that puts GLM 5.2, Claude Opus 4.8, and GPT-5.6 Sol on the quality-per-dollar frontier. Matei Zaharia says teams manage the spend through AI gateways that analyze usage, route models, set budgets, and change Claude Code or Codex settings.

Alibaba's Qwen3.8-Max is live on Venice and OpenRouter while open weights are still described as coming soon. Reports cite a 2.4T-parameter model with strong Vals and vision benchmark results.

Baseten and Together AI added DeepSeek V4 Flash with a 1M-token context window, reasoning-effort controls, and DSpark decoding. ValsAI ranked it the cheapest model above 60 on its index, and Nous promoted a short 90% discount.

New tests showed DeepSeek V4 Flash as cheaper per token and faster on some serving paths. Ramp said it cost 3x more than GPT-5.6 Luna per SWE-Bench task because it used more turns.

DeepSeek released V4 Flash 0731 with weights, a technical report, API access, 1M context, MoE routing, and low token prices. Its cited benchmarks show gains on Artificial Analysis, Terminal-Bench, Frontend Code Arena, and agent tests.

Epoch added significant research problems to FrontierMath Open Problems, bringing the set to 50 unsolved math problems. Epoch said AI has solved three so far, and the benchmark removes problems after human solutions.