Thursday, 8 October 2026

Anthropic has released Claude Haiku 5.5, its small, fast model for high-volume work, and cut the price sharply [1]. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, against $1.00 and $5.00 for Haiku 4.5 [1]. Above 100,000 tokens the price rises to $0.50 and $2.50 [1]. Anthropic says that works out to around 75% less on average, after allowing for a new tokenizer that uses slightly more tokens per task [1]. Simon Willison measured one long prompt at around 1.25x as many tokens as on Haiku 4.5 [2].

Willison notes that, up to 100,000 tokens, the price exactly matches OpenAI's GPT-6 Luna, whose own price rises only to $0.20/$0.75 beyond 272,000 tokens [2]. Haiku 5.5 is the first Haiku with an adjustable effort setting [1]; Willison reports that reasoning cannot be turned off and defaults to medium [2]. Anthropic pitches it for summaries, classification, subagent work, customer support and browser use, and says Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding [1]. It is available on all platforms, including Amazon Bedrock [1][3].

Two other price changes ship with it. Cache reads on Sonnet 5.5 drop 50%, from $0.20 to $0.10 per million tokens, which the company says makes Sonnet 5.5 around 20% cheaper on most agentic work [1]. And this week Max and Team subscribers start getting a monthly API credit: $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team [1]. Willison notes the credits do not roll over [2].

Anthropic has published a new version of its Usage Policy, and it takes effect on November 12 [1]. The company says most changes clarify existing rules and add examples for Claude's longer, more independent work [1]. The Verge notes it is the first change to the policy in over a year [2].

Several sections are rewritten. Rules on fake accounts and fabricated news sites, previously scattered across the elections, fraud, privacy and disinformation sections, are merged into a new section on deceptive campaigns, political or commercial [1]. The election section now targets voter deception and election disruption, and the blanket ban on personalised vote and campaign targeting is removed because it blocked civic work such as nonprofits writing voter information in other languages [1]. The weapons ban now names the software and components that make weapons work, and arming drones and other autonomous vehicles [1]. Tracking people without their consent is prohibited, and Claude cannot be used to decide or recommend who to investigate, arrest or charge [1].

Two additions are new. When Claude is connected to hardware that takes autonomous physical actions and could cause injury, a qualified operator must be able to observe and stop it, and the equipment must hold a safe state if Claude disconnects [1]. The policy also bans "sustained and needless abusive or cruel behavior" toward Claude models; Anthropic says this applies only in extreme cases, not to frustration, pushback, dark creative themes or testing [1][2]. Ending the conversation stays the main enforcement tool [1]. The Verge reports that Anthropic did not say whether user bans could follow, and that the rules may be modified for some government contracts [2].

Google Cloud has announced the Gemini agent, which it calls "a universal agent for work", at its Gemini at Work 2026 event [1]. It takes a task from a single prompt box, plans the work, uses skills and tools, connects to a company's business systems and delivers the result inside documents, inboxes and developer environments, the company says [1]. It picks the model for each job and has built-in cost controls [1].

According to [3], that model choice covers both the Gemini and Claude model families today, with private and open models planned later; the announcement gives no comparative results for quality or cost. Users can call it with @Gemini in Gmail, Docs, Sheets, Slides and Chat, or reach it through Slack, Microsoft 365 and the command line, according to [3]. The Verge reports that it runs in the cloud, so it keeps the same context across connected devices, and that it can work as a "coworker agent" with its own identity and an "@agents.company.com" email address [2]. It can split work across sub-agents [2][3] and, according to [3], run workflows for hours or days.

Most people cannot use it yet. The agent is in private preview for enterprise customers [2][3]. Google says wider availability is coming soon for customers on select Workspace Business and Enterprise plans, according to [3]. No price is public. The Verge notes that Google's consumer Gemini Spark agent, launched in May, sits in a separate interface inside the Gemini app [2].

OpenAI says it has banned accounts tied to two covert influence operations, one from Russia and one from Iran, that used its models to support "false front" entities [1]. It rates the Russian operation Category 5 on the six-point IO Breakout Scale, the first Category 5 operation it has disrupted since it began reporting; the Iranian one reached Category 4 [1]. Both landed content in mainstream media outlets, not just on social media [1].

The Russian operators, nicknamed "Dark Clark" after a fake persona called Mia Clark, appear to have controlled a "research platform" in Latin America called the Social Research Center [1]. OpenAI says the evidence indicates its local employees did not know they were working for a Russian group [1]. The operators used VPNs because OpenAI does not allow access from Russia, and they mainly used ChatGPT to write internal reports rather than campaign content [1]. In those reports they claimed to have used a fake email address to get schools in Lima, Peru, to hold events about Ukraine, then planted stories about them in Peruvian and Polish media [1]. CyberScoop reports the group also released a fake audio clip during protests in Bolivia that prompted an official denial [2].

The Iranian operation ran at least seven fake "journalist" personas that pitched articles about the US-Iran war [1][2]. OpenAI says it found almost 100 articles under those bylines across roughly a dozen outlets, according to [2]. The report does not attribute either operation to a specific government agency, according to [2]. OpenAI says it has shared information on the Russian case with the relevant authorities [1].

OpenAI has withdrawn three manuscripts from its public catalogue of AI-generated mathematics, a day after posting it [1][2]. The repository's history page says a sign error in "Algebraicity of Weil classes on split abelian eightfolds" invalidates a stabilization-trace cancellation argument and a construction that two dependent papers relied on [1]. Those two, on Kuga–Satake correspondences for K3 surfaces and on the rational Hodge conjecture for products of K3 surfaces, were withdrawn as well [1]. The withdrawn papers now carry notices explaining the gap [1].

The same update revised 14 other manuscripts with proof repairs and corrected statements, and changed citations in 13 more [1]. OpenAI also added six formalizations, bringing the share of top-line results with Lean proofs to 300 of 719, about 42% [1]. Lean formalizations let software check a proof's logic [2].

According to [2], the original release contained 722 manuscripts, now grouped into 372 families, produced by an unreleased internal model that was posed roughly 4,000 problems. An OpenAI spokesperson said nearly every result came from a single prompt to a single agent, and that many results are not yet understood by the company's own mathematicians, according to [2]. MIT mathematician Andrew Sutherland said such claims should be treated as unverified until the model is released and the results replicated [2]. OpenAI published no prompts and gave only average compute, short of what an advisory group hosted at the Institute for Advanced Study asked labs to disclose, according to [2].

Goodfire, an interpretability startup, has launched monitors that read an AI model's internal signals while it works, instead of having a second model reread everything it writes [2]. They are available to customers of Baseten, which hosts and runs AI models for other companies [2]. Customers choose which risks to watch, including offensive hacking, chemical and biological weapons misuse and reward hacking, and whether a flag is logged, sent for human review or refused [2].

The design is a cascade. Small classifiers called probes read the model's activations at every step; only when one fires does an LLM judge take a closer look [1][2]. In Goodfire's own write-up, its cyber monitors for the open model Kimi K3 detect about 93% of harmful sessions while interrupting at most 5.5% of benign ones, measured against Claude Opus 5's labels on 2.4K sessions [1]. Goodfire says the cascade cuts average added latency per turn from 160 ms to 2 ms without judge reasoning [1]. It estimates that having Opus 5 judge one million Kimi K3 turns would cost on the order of $200K at OpenRouter rates [1].

TechCrunch reports that in Goodfire's tests on Kimi K3, monitoring about 1,500 sessions cost roughly $51, against $233 for a cheaper AI model checking every step and about $10,000 for a top-tier one [2]. These are the company's own figures [2]. TechCrunch reports that the pitch is aimed at open models, which do not come with the kind of monitoring that closed labs run on their own systems [2].