Gemini 4 Argon beats Claude and GPT-6 Astra on 12 of 18 tests but remains reserved for cyber defenders. Its strengths, weaknesses, and why.
OpenAI halves its prices, two days after xAI and Anthropic. Sol improves at coding, Luna declines, and the hallucination rate hides a sleight of hand.
Opus 5.5 is 40% cheaper than Opus 5 and outperforms Fable 5.1 on four tests. The two rows the manufacturer puts at the bottom of its table.
Grok 4.7 halves the bill, MiMo v2.6 can be downloaded under the MIT license. The figures, the exam it fails, and what it would take to run it.
Sakana AI sells a model that chooses which of its models answers. It beats GPT-6 Astra and Fable 5.1 in its own tests.
98.6% on the hardest test, 61 out of 100 from the independent judge: where the gap comes from, and what GPT-6 Astra really changes.
Corrected comparison of AI coding tools: what to choose for your profile, price per million tokens, and actual limits.
ByteDance is training a model with up to 10,000 billion parameters. Why this figure says almost nothing, and what the Doubao story really hides.
DeepSeek reports 82.7 on Terminal-Bench versus Opus 4.8’s 85.0. The official leaderboard says 78.9. What the evaluation harness really changes.
Ce site utilise des cookies pour améliorer votre expérience. En continuant à naviguer sur ce site, vous acceptez notre utilisation des cookies. Accepter Refuser
