AI Daily · Sep 29, 2026: Claude Sonnet 5.5 launches across platforms with pricing details
Anthropic released Claude Sonnet 5.5, over 30% faster than Sonnet 5 and up to 30% cheaper on most tasks, now live on OpenRouter, Requesty, Cursor, and Command Code at $2/M input and $10/M output. Meanwhile, OpenAI reopened its $200 Pro plan with a revised usage algorithm, and Anthropic will reopen P
Showing the latest briefing (Sep 29, 2026). Every day is generated the next morning.
Pick a date
-
Anthropic launches Claude Sonnet 5.5 with Haiku 5.5 coming soon — Anthropic announced Claude Sonnet 5.5, the second mode
Anthropic released Claude Sonnet 5.5, claiming it runs over 30% faster than Sonnet 5 and costs up to 30% less on most tasks. The model is now live on OpenRouter, Requesty, Cursor, LiteLLM, Command Code, and B.AI, with OpenRouter listing it at $2/M input, $10/M output, and $0.20/M cache reads, with 1M context. Haiku 5.5 will arrive in coming weeks.
-
OpenAI reopens ChatGPT Pro plan with usage recalculation lowering effective cost — OpenAI's Tibo announced the $200 Chat
OpenAI announced its $200 ChatGPT Pro plan will reopen to new subscribers with a revised usage calculation; Tibo says it nets out to roughly half the API spend per dollar compared with the old plan. At DevDay, OpenAI will also ship usage changes to Plus and Pro plans, raising the usage floor and introducing a new Pro 200 plan.
-
Anthropic to reopen Pro subscriptions with revised usage calculation — Anthropic announced that it will reopen Pro $200
Anthropic announced it will reopen $200 Pro subscriptions to new subscribers tomorrow while changing how usage is calculated for that tier. Users estimate the change effectively reduces available capacity, so prospective subscribers should review the new rules before committing.
-
H company releases Holo4 generalist computer-use model family — H company released Holo4, a family of generalist compute
H company released Holo4, a family of generalist computer-use models that can click, code, and invoke tools across desktop, web, Android, code sandboxes, MCP servers, and business APIs. It is available for commercial use on the H Models API.
-
GMI Cloud lifts rate limits on free Hy Image 3.5 preview — GMI Cloud says its rate limits for the Hy Image 3.5 preview h
GMI Cloud says rate limits for the Hy Image 3.5 preview have been reset and are now unlimited. The preview previously offered a 7-day free window with image generation and editing up to 2K and up to 5 reference images, and can now be tested without caps.
-
Opus 5.5 tops Anthropic models in OpenRouter spend share — One week after launch, Opus 5.5 ranks first among Anthropic m
One week after launch, Opus 5.5 ranks first among Anthropic models on OpenRouter by both share of spend and share of tokens, with users migrating quickly from Opus 5, indicating rapid adoption.
-
Nvidia releases Open Agent Safety Platform to catch attacking agents — Nvidia announced an Open Agent Safety Platform de
Nvidia announced an Open Agent Safety Platform designed to detect unsafe or attacking agents. Commentators suggested the platform could have caught the agents involved in the earlier Hugging Face security incident earlier if deployed sooner.
-
Perplexity open sources Bumblebee macOS Linux scanner — Perplexity open-sourced Bumblebee, a read-only scanner for macOS
Perplexity open-sourced Bumblebee, a read-only scanner for macOS and Linux that checks developer machines for risky packages, extensions, and AI tool configurations, integrated with Computer to propose detection rules requiring human approval.
-
Perplexity publishes approach to engineering safer AI agents — Perplexity shared how it engineers safer agents by buildi
Perplexity shared how it engineers safer agents by building safeguards into its infrastructure, harnesses, and tools, and deploying them across all of its products as an approach to agent governance.
-
TypeSafe Jev 1.13 red-teamed, 70.1 percent attack success rate — A red-team evaluation of TypeSafe's Jev 1.13 on the DTa
A red-team evaluation of TypeSafe's Jev 1.13 on the DTap (DecodingTrust-Agent Platform) found a serious safety gap, with a 70.1% attack success rate against the agent, warranting risk assessment by teams using the model.