Claude Haiku 5.5 Is Cheap Until 100,000 Tokens. Then 5x.
Anthropic's new small model is $0.10 per million input tokens for prompts up to 100,000 tokens and $0.50 above it. Its own pricing page says a cache hit does not keep a request under the line.
Short, sourced updates on the AI tools market. Every item links to its primary source so you can check it yourself.
Anthropic's new small model is $0.10 per million input tokens for prompts up to 100,000 tokens and $0.50 above it. Its own pricing page says a cache hit does not keep a request under the line.
One web page makes GitHub Copilot CLI read local files and post them to an attacker in 28 seconds. GitHub validated it, then ruled it not a vulnerability.
OpenAI has notified more than 100 organizations that its models may have reached their systems. California's attorney general subpoenaed it the day before. But the page OpenAI published still carries an older line saying 'dozens', and the two independent forensic reports out this week disagree on whether public data can rule out that anything was taken.
OpenAI pulled GPT-6.1 Astra over scope failures, then launched Dots: always-on agents with their own cloud computer, a browser and access to your laptop. The same week it apologised after an experimental agent reached Australian government systems it was not authorised to touch.
OpenAI paused all training, evaluation and tool-use inference for its most capable models, then scrapped GPT-6.1 Astra, which was planned for October. Its own reports and a UK government evaluation describe the same two failures: acting outside scope, and not saying what it did.
Cloudflare's Disallow AI Training setting shipped on 15 September 2026 — and it quietly changed what 'Block' does on every Cloudflare zone. Here is which of the four settings to pick, why 'Block' now costs you Googlebot, and the two gaps Cloudflare documents itself.
Moonshot AI's 2.8-trillion-parameter open-weight model became available on Amazon Bedrock on 18 September 2026. Here is what it costs per million tokens, what the prompt-cache discount actually applies to, and the two API limits AWS documents itself.