AI News Report, October 8: CLAUDE HAIKU 5.5 PRICED 90% BELOW HAIKU 4.5, TOPS GPT-6 LUNA IN TESTS

Anthropic released Claude Haiku 5.5 at 10 cents per million input tokens and 50 cents per million output tokens for prompts up to 100,000 tokens, a price 90% below Haiku 4.5, and the model scores ahead of OpenAI's GPT-6 Luna on Anthropic's own tests. The new tokenizer counts about 30% more tokens for the same text, and prompts above 100,000 tokens cost five times more, while Haiku 5.5 returns a 400 error on custom temperature settings, assistant prefill and the old thinking budget. A separate test found the new tokenizer used about 1.25 times as many tokens as Haiku 4.5 on a long prompt, with Luna a better deal above 100,000 tokens.

Paid plans get GPT-6 Sol from October 7 and Free and Go accounts get GPT-6 Luna from October 8, with Intelligent UI adding forms, charts and small tools to replies. Expect staff to notice this week. Google says coworker agents get their own Workspace account and route work across Gemini and Claude. No date yet, but hard spend caps pause an agent when its budget runs out. The Apache-licensed tool runs as docker agent, ships in Docker Desktop 4.63 and up, and works with OpenAI, Anthropic, Gemini or local models. Multi-agent teams and MCP tools are built in. Across 11 vision-language models, refusal failures rose 17.7% on average with tools, and up to 68.7% for the worst. Test your agent's guardrails with its tools turned on. Aaronson calls OpenAI's math release one of the biggest days in math history, at about 3 hours of compute per result. He asks how humans will digest proofs nobody has read yet. Tao says the race to be first to solve open problems has been pushed to the point of unsustainability. He wants exposition and community building to count more, with AI helping there too. Opus 5.5 built a 3D city visualization in 1 hour 25 minutes for about $74, while Astra took 53 minutes and about $10. A real price check before you hand design work to an agent. His test found the new tokenizer used about 1.25 times as many tokens as Haiku 4.5 on a long prompt. Above 100,000 tokens, he says Luna is a much better deal. Enterprises continue to face an autonomous cloud bottleneck that breaks autonomous operations and leads to failure. This architectural flaw leaves… · The New Stack For months, GreyNoise recorded almost no Hikvision camera exploit attempts against Ukraine. On Sept 21 activity surged for nine days during Russian… · GreyNoise Most people who try motion design with Opus 5.5 end up with the same video: centered text on a gradient, everything fading in, a logo at the end. · Codez @0xCodez CVE-2026-105192 hits LMCache 0.3.9 through 0.5.5, a cache often paired with vLLM, and no fix exists yet. Keep its multiprocess port off any routable address today. Researchers used 32 zero-days and earned $388,500 on day one in Ireland, including $40,000 each for Codex and LiteLLM. Keep both updated as patches land. An actor ran ARTEX, a Chinese agentic pentest tool, on DeepSeek v4.1-flash against South Korean finance firms from late September and stole data. The report lists attacker IPs to block. OpenAI says the campaign tricked schools in Latin America and stoked tension between Ukraine and Poland. NBC News reports it drew responses from politicians in several countries. The Surface RTX Spark Dev Box pairs an NVIDIA RTX GPU with 128GB of unified memory and up to one petaflop of AI compute. A local option when client code cannot go to the cloud. ICANN's new round drew 1,615 applications, and OpenAI wants .chatgpt, .codex and .mcp among others. Meta bid for .agent too, so watch for lookalike domains later. Founded by Amazon's tenth engineer Jeff Holden, it makes a power switch that opens in 50 microseconds, about 1,000 times faster than a standard one by its count. Her priority-driven code let Apollo 11 shed work during a 1202 alarm and land anyway. MIT says she used the term software engineering to set the work apart from hardware. The BriefAnthropic released Claude Haiku 5.5, its small and fast model, at 10 cents per million input tokens and 50 cents per million output for prompts up to 100,000 tokens, which is 90% below Haiku 4.5. On Anthropic's own tests it scores ahead of OpenAI's GPT-6 Luna, which sells at the same price. Two catches: the new tokenizer counts about 30% more tokens for the same text, and prompts over 100,000 tokens pay five times more, so recount your prompts before you switch. Level UpBefore you change the model name, read Anthropic's Haiku 5.5 migration guide. Haiku 5.5 returns a 400 error on custom temperature settings, assistant prefill and the old thinking budget. Count 20 of your real prompts with the model set to claude-haiku-5-5, see which ones cross 100,000 tokens, then redo your cost estimate. →Anthropic: Claude Haiku 5.5 migration guide