Try
Useful tools & techniquesMicrosoft-Decision-1 adds low-cost structured scoring
Microsoft-Decision-1 is available through Microsoft Foundry and OpenRouter for routing, classification, grading, and agent control. Microsoft lists input at $0.042 per million tokens and free output. The model returns probabilities for fixed choices; Microsoft reports a 1.3% decision-flip rate under eight input perturbations and claims it is 35 times faster than GPT-6 Sol at median latency.
Test it as a probability-scoring or agent-gating step; benchmark vendor claims on your own workload.
Celeris offers a fast hosted decision API
Celeris-1-decision accepts text, structured data, or images and returns probabilities for typed choices; explanations are optional. Celeris reports 80% accuracy on JevBench, 67 ms median latency, and $0.04 per million input tokens, with output free. It is hosted only, with no weights or fine-tuning; its request format follows the System One convention.
Point a compatible decision client at Celeris and test calibration as well as top-choice accuracy.
OpenAI math results become executable Lean tasks
FineEnvs packaged 369 verified OpenAI math proofs as Harbor environments covering 214 result families across 17 fields. Agents must replace Lean 4 placeholders with complete proofs of the exact formalized statements; a separate Lean checker grades submissions. Run tasks with Harbor 0.24+, Daytona, and an agent credential; each task requests four CPUs, 8 GB memory, and up to four hours.
Use the set to evaluate theorem-proving agents, but expect most open problems to score zero.
Know
What changedMeta opens Muse Wonder; Wonder Full stays in Muse Code
Meta announced Muse Wonder for its API and Muse Code, while Wonder Full is limited to Muse Code users. Artificial Analysis reports scores of 52 and 56, respectively; Wonder costs $4/$20 per million input/output tokens, with a 1M-token context. Meta’s subscriptions offer four usage tiers, default to content-contributing mode, and let users opt out. A post citing Sensor Tower figures puts Muse retention at 22.5% on day 14, versus ChatGPT’s 26.8%.
Check model access and contribution settings before routing private code through a Muse Code subscription.
Cursor bundles Grok and X into opt-in plans
New plans combine Cursor, Grok, and X, with weekly shared usage pools: Plus is $8/month, Premium $30, Super $100, and Ultra $200. Existing subscribers keep current terms unless they opt in; new sign-ups use the new plans. Third-party models draw credits at provider list price, so compare your model mix before switching.
Audit third-party model spend first; the bundled pool may not save heavy users of other providers.
Mistral Large 4 preview is live; weights due this month
Mistral Large 4’s public API preview is available in Mistral Studio; the company says weights are planned by month-end. It is a multimodal 1T-parameter model with 52B active parameters. Mistral reports 61.7% on DeepSWE v1.1 and 82% on a vulnerability-reproduction-and-patching test; its published API pricing is $1.36 input and $4.18 output per million tokens.
Try the preview for coding or security workflows, but treat vendor benchmark results as claims pending independent testing.
Visual-reasoning benchmark exposes persistent model gaps
Humanity’s Sixth Sense tests implicit visual reasoning across 522 image and video tasks; the reported human score is 93.1%, versus 53.6% for GPT-6 Astra. Tool-using agents improved the best result to 59.3% on a subset, but did not close the gap.
Treat visual judgments about scene context or physical relationships as review-required.