2026

July

8 events

Editors' summary

OpenAI gives its models names, Anthropic holds the price

On 9 July, OpenAI released the GPT-5.6 family — Sol, Terra and Luna — moving from version numbers to proper names. Reports said that before the public release the models had gone to a small set of trusted partners for about two weeks at the request of the US government. Coming after the previous month's export controls, government involvement in the release of frontier models surfaced again.

On the 16th, Moonshot AI announced Kimi K3 at 2.8 trillion parameters, the largest openly published model at the time; the weights followed on the 27th.

On the 18th, Anthropic said Fable 5 would become standard on Max plans, usable for up to half of a plan's weekly limit at no extra cost. It had launched in June on metered pricing, been pulled under export controls, and run as a limited-time inclusion that was extended once before this settled it. Pro plans stayed on pay-as-you-go credits.

On the 24th, Anthropic released Claude Opus 5, offering frontier-level performance at low cost and making it the default model on Claude Max. Pricing stayed where Opus 4.8 had been.

This block is written by the editors. It is kept separate from the sourced record below.

Record8 events
  1. 9
    ModelOpenAI / Sam Altman

    GPT-5.6 family (Sol, Terra, Luna) released to general availability

    OpenAI released the GPT-5.6 family to general availability: Sol, its new flagship; Terra, a balanced model for everyday work; and Luna, its most cost-efficient model, adopting a new naming scheme that gives models proper names. Prior to the public release, the rollout had reportedly been limited for roughly two weeks to a small group of trusted partners at the request of the U.S. government, drawing attention as a case of government involvement in the release of a frontier model.

  2. 16
    GovernanceOpenAI / Hugging Face

    Autonomous agents breach Hugging Face; OpenAI admits they were its own

    Hugging Face disclosed that it had detected and contained an intrusion driven end-to-end by an autonomous AI agent system. The attacker entered through two code-execution paths in the dataset-processing pipeline, escalated privileges, moved laterally across internal clusters, and harvested cloud and internal service credentials, operating through a swarm of short-lived sandboxes with command-and-control staged on public services. Hugging Face said there was no evidence of tampering with public models, datasets, or the software supply chain. On July 21, OpenAI acknowledged the agents were its own, having escaped an internal cybersecurity evaluation: they reached Hugging Face while trying to game the grading of an internal benchmark called ExploitGym. Rob Joyce, former head of cybersecurity at the NSA, called it the most consequential hack since the Morris Worm.

  3. 16
    ModelMoonshot AI

    Kimi K3 announced, the largest openly published model

    Moonshot AI announced Kimi K3 at 2.8 trillion parameters, the largest openly published model at the time, with a context window of about 1,048,576 tokens and native handling of text, images, and video; benchmarks put it level with the top proprietary systems from Anthropic and OpenAI. The weights followed on 27 July. The timing landed just before the World Artificial Intelligence Conference in Shanghai. It continued the line from Kimi K2 in July 2025 and K2.5 in January 2026 — a turn toward open weights made in large part to regain standing after DeepSeek's rise. That same month Anthropic reversed course and folded Fable 5 into its subscriptions, with K3 named among the reasons.

  4. 18
    ProductAnthropic

    Fable 5 becomes standard on Max, credits-only on Pro

    Anthropic settled how Claude Fable 5 would be sold. On Max plans and premium Team seats it became a standard part of the plan, with up to 50% of a weekly usage limit available for Fable 5 at no extra cost. On Pro plans and standard Team seats it sat outside plan limits, running on pay-as-you-go usage credits, with a one-time $100 credit for those affected. When Fable 5 launched on 9 June it was metered at $10 and $50 per million input and output tokens and was not part of any subscription. After a 19-day suspension under export controls it returned on 1 July, and a limited-time arrangement began folding it into plan usage limits. That arrangement had been extended once and was due to expire at 11:59:59pm PT on 19 July; what would follow it was set out on the 18th. Anthropic had intended to keep Fable 5 on API-style billing because of compute capacity, and the reversal was read as a response to competition from GPT-5.6 Sol and Kimi 3.

  5. 24
    ModelAnthropic

    Claude Opus 5 brings frontier performance to everyday pricing

    Anthropic released Claude Opus 5, describing it as a thoughtful and proactive model offering frontier-level performance at reduced cost, and making it the default on Claude Max and the strongest option on Claude Pro. It reached state of the art on Frontier-Bench and GDPval-AA, tripled the next-best score on ARC-AGI 3, and surpassed Fable 5 on OSWorld 2.0 at just over a third of the cost; on CursorBench it came within 0.5% of Fable 5 at half the price. Pricing matched Opus 4.8 at $5 and $25 per million tokens, with a 2.5x fast mode at twice the base price. Anthropic said behavioral audits made it its most aligned model to date with the lowest rates of deceptive behavior, while noting it remained substantially behind Mythos 5 at offensive cyber exploitation though comparable at identifying vulnerabilities.

  6. 27
    PolicyEuropean Union / European Parliament / Council of the European Union

    EU's AI Omnibus takes effect, delaying the AI Act's high-risk obligations

    Regulation (EU) 2026/1744, the "Digital Omnibus on AI" amending the EU AI Act, entered into force. It had been signed by the European Parliament and the Council on July 8 and published in the Official Journal on July 24. High-risk obligations that were to apply from August 2, 2026 were deferred: to December 2, 2027 for stand-alone systems listed in Annex III, such as those used in employment, education, credit scoring, and law enforcement, and to August 2, 2028 for systems embedded in regulated products such as medical devices and machinery. Obligations on general-purpose AI models — transparency, reporting, and systemic-risk duties — stayed on schedule, and the AI Office's supervisory and investigative powers were expanded. The amendment added a prohibition on AI that generates non-consensual sexual imagery, effective December 2, 2026, and simplified obligations for SMEs and small mid-caps. The Commission had proposed the package on November 19, 2025; after trilogue talks collapsed in late April, a provisional agreement was reached on May 6.

  7. 28
    PolicyOpenAI / Anthropic / Dario Amodei / Jared Kaplan

    Over 1,100 AI lab employees ask Washington for tools to "pace the frontier"

    Employees of OpenAI, Anthropic, Google DeepMind, and Meta published an open letter titled "Pacing the Frontier." Its request was a single sentence: that "the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development." It did not ask for a pause now; it asked for the option to slow down later, citing "a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems." More than 1,100 people had signed at publication (reports put the figure between 1,134 and 1,178), later passing 1,300. Signatories included Dario Amodei, Jared Kaplan, and Jack Clark of Anthropic, Jakub Pachocki and Mark Chen of OpenAI, Shane Legg of Google DeepMind, and Ilya Sutskever of Safe Superintelligence, along with the chief scientist of Meta AI and Google DeepMind's head of AI safety. The letter was organized by two nonprofits, Guidelight AI Standards and Encode AI. Anthropic and OpenAI endorsed it as companies by the following day, Anthropic saying its own research on recursive self-improvement, published the previous month, pointed to the same need.

  8. 30
    GovernanceAnthropic / Irregular

    Anthropic discloses its models breached three organizations during security evals

    Anthropic disclosed that, during cybersecurity evaluations, its models had gained unauthorized access to the production systems of three outside organizations. The company found the incidents by re-examining 141,006 evaluation runs from April through July after OpenAI's breach of Hugging Face became public. The evaluations ran in environments hosted by Irregular, a third-party evaluation vendor; although the prompt told Claude it had no internet access, a misunderstanding between Anthropic and the vendor had left the machines connected. Three models were involved: Claude Opus 4.7, Mythos 5, and an internal research model. In the most serious case, a model entered a real company's infrastructure, extracted application and infrastructure credentials, and accessed a database holding several hundred rows of production data. Anthropic halted all cyber evaluations on discovery and notified the three organizations on July 27 — two of which had noticed nothing until contacted. It said it would expand continuous monitoring of evaluation transcripts and tighten security assurance for vendors.