All Issues
Jun 22 - Jul 12, 2026

AI Weekly: GPT-5.6 Goes Live, Meta Pivots to Closed API, Fable 5 Costs Extra

Models & Releases

5 stories

GPT-Live: Full-Duplex Voice That Listens and Talks

  • OpenAI launched GPT-Live, a new full-duplex voice architecture that listens and speaks simultaneously, enabling real-time back-and-forth including active listening cues like ‘mhmm’ without rigid turn-taking.
  • GPT-Live delegates complex queries to a background frontier model (currently GPT-5.5), keeping the conversation flowing while deeper work completes—decoupling real-time interaction from heavy reasoning.
  • Two versions—GPT-Live-1 and GPT-Live-1 mini—are rolling out globally to ChatGPT users now; API access is coming soon and developers can sign up for early access.

Claude Sonnet 5: Near-Opus Performance for Agents

  • Anthropic launched Claude Sonnet 5 on June 30 as its most agentic Sonnet yet, closing the performance gap with Opus 4.8 on BrowseComp and OSWorld-Verified while offering better cost efficiency.
  • It is now the default model for Free and Pro plans, available across Claude Code and the Claude Platform at introductory pricing of $2/$10 per million tokens through August 31, 2026.
  • Early partners reported it completes multi-step tasks—Salesforce tier updates plus launch emails end-to-end, complex GitHub PRs—where previous Sonnet models would stop short, and proactively checks its own output.

Gemini 3.5 Flash Gets Native Computer Use

  • Google DeepMind integrated computer use natively into Gemini 3.5 Flash on June 24, enabling agents that can see, reason, and act across browser, mobile, and desktop environments without a separate model.
  • Targeted adversarial training mitigates prompt-injection risks; two optional enterprise safeguards allow requiring confirmation for irreversible actions and auto-stopping on indirect injection detection.
  • Available via the Gemini API and Gemini Enterprise Agent Platform; Browserbase is providing a public demo environment for developers to test immediately.

Meta Muse Spark 1.1 and the Meta Model API

  • Meta Superintelligence Labs released Muse Spark 1.1 on July 9 — a multimodal reasoning model with a 1M-token context window, zero-shot MCP generalization, and major gains in tool use, coding, and computer use — priced at $1.25/$4.25 per million input/output tokens.
  • Alongside it, Meta opened the Meta Model API (public preview, US-only), a closed hosted API that marks a historic strategic pivot away from Meta’s open-weights Llama philosophy toward metered, proprietary inference.
  • Meta also launched Muse Image (advanced image generation with multi-reference composition and AI effects in Instagram Stories) and previewed Muse Video — all developed by Meta Superintelligence Labs.

People & Business

2 stories

Fable 5 Goes Pay-Per-Token Starting July 12

  • Anthropic is moving Claude Fable 5 behind usage-based billing starting July 12 at 11:59 PM PT — subscribers on $20, $100, and $200/month plans will owe additional fees at $10/million input and $50/million output tokens, matching API rates.
  • This appears to be the first time a frontier AI lab has gated a consumer chatbot model behind per-token billing on top of a subscription, ending the era of unlimited frontier AI access for a flat monthly fee.
  • A million tokens equals roughly 750,000 words (longer than all of Lord of the Rings), so the fee affects power users and heavy chain-of-thought sessions most; lighter use stays within normal subscription budgets.

Policy & Ethics

2 stories

OpenAI Bio Bug Bounty Expands, Raises Reward to $50K

  • OpenAI is evolving its GPT-5.5 Bio Bug Bounty into an ongoing private program — the OpenAI Bio Bounty Program — now focused on universal jailbreaks against GPT-5.6 and future frontier models.
  • The reward for a successful universal jailbreak has been raised from $25,000 to $50,000 for both GPT-5.6 and GPT-5.5; testing for GPT-5.5 ends July 27, after which only GPT-5.6 is in scope.
  • The program is invite-only via rolling application; participants must sign an NDA, and the expansion signals OpenAI is treating biosafety red-teaming as a permanent, structured part of its safety infrastructure.

Products & Hardware

3 stories

Claude Tag: @Claude Joins Your Slack as a Teammate

  • Anthropic launched Claude Tag on June 23 for Slack — Claude joins channels as a persistent team member with shared context, responding when @tagged and proactively surfacing relevant updates in ambient mode.
  • It is multiplayer (one Claude per channel; anyone can hand off mid-conversation), learns from channel history, and can schedule tasks to run autonomously over hours or days without prompting.
  • 65% of Anthropic’s own product team code is now generated by an internal version of Claude Tag; beta is live for Claude Enterprise and Team customers, with broader platform expansion planned.

OpenAI and Broadcom Unveil Jalapeño Inference Chip

  • OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first Intelligence Processor — a blank-slate LLM inference accelerator designed around OpenAI’s deep understanding of LLM kernels, memory movement, and serving patterns.
  • Engineered samples are already running GPT-5.3-Codex-Spark at production frequency; early testing shows performance-per-watt substantially better than current SOTA, with a full technical report due in coming months.
  • The chip is the first in a multi-generation compute platform built with Broadcom and Celestica for gigawatt-scale data center deployment with Microsoft and partners beginning in 2026.

Research & Resources

3 stories

Claude Science: AI Workbench for Computational Biology

  • Anthropic launched Claude Science on June 30 as a unified research environment that integrates PubMed, Jupyter, R, cluster terminals, and HPC login nodes, with 60+ curated skills and connectors for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
  • Every output carries an auditable history — figures include the exact code, environment, and message history that created them; a reviewer agent independently checks citations and calculations.
  • Claude Science is in beta for Pro/Max/Team/Enterprise users; Anthropic is also offering AI for Science grants (up to 50 projects, up to $30K credits each) with applications open through July 15, 2026.

OpenAI Audits SWE-Bench Pro: ~30% of Tasks Are Broken

  • OpenAI published a systematic audit of SWE-Bench Pro—one of the most widely used coding benchmarks—finding evidence of breaking issues in approximately 30% of its 731 tasks, including overly strict tests, underspecified prompts, low-coverage tests, and misleading prompts that contradict what hidden tests require.
  • The audit used a data quality pipeline flagging 200 tasks (27.4%), confirmed by human annotation identifying 249 (34.1%); findings were reviewed by five experienced software engineers with disagreements escalated for independent review.
  • OpenAI advises model developers to treat SWE-Bench Pro results with caution and highlights the growing utility of agents for scalable data quality checks — extending the SWE-bench Verified concerns raised earlier.