Section
Artificial intelligence
Model capability, evaluation integrity, and what inference actually costs.
32 stories · 0 editors
Topics
Sort: Newest ▾
News
4 min
OpenAI's GPT-6 Sol and GPT-6 Luna arrived on September 22 at half the promotional price of their GPT-5.6 namesakes. The rate card shows the Sol name now sits one rung lower than it did.
News
4 min
Anthropic released Claude Sonnet 5.5 on September 28 at Sonnet 5's prices. Its migration docs list five requests that now fail, and one of them collides with the model's new cyber fallback.
15 Sep · Prelo · 6 min
GPT-6 Astra drops temperature and none effort, and its misalignment monitor can end an agent run the API cannot resume.
09 Sep · Prelo · 4 min
ant apply brings a plan-and-approve loop and a claude-lock.json to Claude API resources.
09 Sep · Prelo · 4 min
Gemini 3.8 Flash is GA; 3.8 Flash Cyber is gated behind Google's Fairwind Program.
04 Sep · Prelo · 5 min
Friday's unannounced weekly-limit reset, the Reddit threads that timed it, the Astra coincidence, and an undocumented /limit-reset in the binary.
01 Sep · Prelo · 3 min
Claude Code's limits: reset today, up 25% on September 14, and down 17% from where they are now.
01 Sep · Prelo · 13 min
Fable 5.1 held its price, cut cache reads to $0.25, and broke three requests that worked on Fable 5.
01 Sep · Prelo · 3 min
OpenAI's Assistants API is gone; unexported Threads history went with it.
01 Sep · Prelo · 3 min
Anthropic ships a browser use toolset and takes computer use out of beta.
30 Aug · Prelo · 9 min
Anthropic's J-space paper, and the ablation that separates what Claude can report from what Claude can do.
25 Aug · Prelo · 5 min
Sonnet 5 keeps $2/$10; Gemini 3.7 Flash's intro rate expires 2026-12-31. Read the tokenizer note.
25 Aug · Prelo · 6 min
httpx2, no more temperature on Messages, and a Bedrock default that now raises instead of picking us-east-1.
18 Aug · Prelo · 5 min
Anthropic details how Claude's text watermark works — and where it does not: code, short passages, and factual prose.
14 Aug · Prelo · 9 min
Anthropic's red team on how agent swarms fail — conformity, collusion, sabotage, and the coordination that intelligence doesn't buy.
11 Aug · Prelo · 7 min
Claude Enterprise can now route every governed prompt to an org-run security server for a verdict before inference. Read the failure modes first.
08 Aug · Prelo · 4 min
Cross-session messaging ships in Claude Code, with a security posture worth reading before the feature.
04 Aug · Prelo · 4 min
A compliance-day statement that addresses safety and provenance, and leaves the Code's copyright chapter unmentioned.
04 Aug · Prelo · 3 min
Gemini 3.6 Flash and later ignore temperature/top_p/top_k; future models will return HTTP 400.
01 Aug · Prelo · 3 min
Claude refusals arrive as HTTP 200 with an empty content array. Clients that index content[0] crash; stop_reason is the field to check first.
31 Jul · Prelo · 3 min
The prompt-cache minimum drops to 512 tokens on Opus 5, Fable 5 and Mythos 5. Prompts written off as uncacheable now cache — here is how to check yours.
30 Jul · Prelo · 3 min
Fable 5 requires 30-day data retention — zero-data-retention orgs get a 400 on every request. The workaround, the scope, and the procurement consequence.
Older in AI →