AI Bulletin · week of June 29 – July 3, 2026
Recap of the week of June 29 – July 3, 2026 in AI. Six news blocks and a closing framework with two data series on AI and employment.
This week’s bulletin gathers six news blocks on AI applied to work, from 29 June to 3 July 2026, plus a closing framework with two data series on AI and employment. The week was marked by the return of Claude Fable 5 and Mythos 5 after export controls were lifted, and by a dense run of model launches.
1. Fable 5 and Mythos 5 return: Commerce lifts the export controls
Anthropic announced on 1 July that the US Department of Commerce has lifted the export controls on Claude Fable 5 and Mythos 5, and that it would start restoring access the following day. The announcement closes the episode opened in mid-June with the withdrawal of both models, and confirms the return signals that circulated the previous week.
A day later, the company published the redeployment terms: Fable 5 is included, until 7 July, in at most 50% of weekly usage limits, and will move to usage credits afterwards. Mythos 5 has been restored for some organisations in the United States, and Anthropic says it is coordinating with the US government to expand access to the full set of domestic and international partners in the Glasswing program.
On the same government-industry front, CNBC reported that OpenAI has proposed giving the Trump administration a 5% stake to ease pressure from Washington. The proposal is part of a larger arrangement under which the US government would hold 5% of each of the country’s leading AI developers through a sovereign wealth fund vehicle.
Sources: Anthropic statement via TLDR AI, 1 July — “The Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5”; Anthropic, 2 July — “Redeploying Fable 5”; CNBC, 2 July — “OpenAI proposes 5% stake to Trump administration to ease Washington pressure”.
2. GPT-5.6 Preview, Grok 4.5 and Watermelon: a frontier-model week
OpenAI introduced GPT-5.6 Preview, a family of models named Sol, Terra and Luna, with Sol positioned as the lead model. The system card describes strengthened cyber and bio safety testing, new safeguards, and a limited preview before broader availability.
Elon Musk said Grok 4.5 is in private beta at SpaceX and Tesla. According to his statements, the model is based on a 1.5-trillion-parameter V9 foundation model with Cursor data added during supplemental training, and early evaluations place it near or above Opus, with reinforcement learning still in progress.
At Meta, superintelligence chief Alexandr Wang said the company’s upcoming model, codenamed Watermelon, has caught up with OpenAI’s GPT-5.5 on the most closely followed benchmarks. The model is still in training and reportedly uses an order of magnitude more compute than Muse Spark. Meta has not given a timeline.
Sources: OpenAI, 29 June — “GPT-5.6 Sol, Terra, and Luna”; report via TLDR AI, 29 June — “Musk says Grok 4.5 entered private beta”; Let’s Data Science, 3 July — “Meta’s Watermelon matches GPT-5.5 benchmarks”.
3. A loaded week at Anthropic: Sonnet 5, Claude Science, drugs and chips
Anthropic introduced Claude Sonnet 5, a lower-cost Sonnet model with stronger agentic performance in planning, tool use, coding and knowledge work. According to the company, its capabilities approach Opus 4.8 and improve substantially over Sonnet 4.6.
It also announced Claude Science, an AI workbench app available in beta for Pro, Max, Team and Enterprise users on macOS and Linux. The product integrates fragmented scientific tools into a single environment and natively renders 3D protein structures, genome browser tracks and chemical structures. In parallel, CNBC reported that Anthropic is starting an internal drug discovery program to develop AI tools aimed at drugmakers, focused on neglected diseases that traditional biopharmaceutical companies would not consider attractive targets.
On the corporate side, TechCrunch reported that Anthropic has discussed a custom AI chip collaboration with Samsung to diversify its compute; the company said chips from Google, Amazon and Nvidia remain central to its hardware strategy. And The Next Web picked up the confusion at Salesforce over the promotion of Claude Tag inside Slack: the company expects to spend $300 million on Anthropic tokens this year and holds roughly a 1% stake in the company.
Sources: Anthropic, 1 July — “Claude Sonnet 5” and “Claude Science, an AI workbench for scientists”; CNBC, 30 June — “Anthropic launches AI drug discovery program, joining tech giants in betting on healthcare”; TechCrunch, 2 July — “Anthropic is discussing a new custom chip with Samsung”; The Next Web, 30 June — “Salesforce employees are confused about why the company is promoting a competitor inside Slack”.
4. Google and Meta: generative image, compute capacity and a cloud of Meta’s own
Google launched Nano Banana 2 Lite, which it describes as its fastest and most cost-efficient Gemini Image model, alongside Gemini Omni Flash for video generation and conversational editing. Both are available through AI Studio, the Gemini API, and the company’s consumer and enterprise products. In addition, the personalised image generation powered by Nano Banana became free for users in the United States: the feature generates images from what the model knows about the user’s preferences, within the opt-in Personal Intelligence feature.
On the infrastructure front, CNBC reported, citing the Financial Times, that Google has limited Meta’s access to Gemini capacity after Meta requested more compute than Google could provide; the shortfall is said to have delayed some of Meta’s internal AI projects. Days later it was reported that Meta is preparing a cloud business to sell its surplus AI computing capacity and hosted models to external developers, in direct competition with Amazon Web Services, Microsoft Azure and Google Cloud.
Sources: Google, 1 July — “Nano Banana 2 Lite”; TechCrunch, 29 June — “Gemini’s personalized AI image generation is now free for US users”; CNBC (citing the Financial Times), 28 June — “Google limits Meta’s use of its Gemini AI models”; report via TLDR AI, 2 July — “Meta is planning a cloud business to sell AI computing power”.
5. Coding agents: Devin Fusion, Security Swarm, Cursor for iOS and ZCode
Cognition introduced Devin Fusion, a multi-model harness that mixes frontier models with lower-cost ones and cuts expenses by 35% on the FrontierCode benchmark while maintaining performance. The system uses a dual-agent architecture, a main agent and a sidekick, to route dynamically between models; with Fable 5 integration, the cost reduction reaches 41%. The same week, the company published Devin Security Swarm, a system for finding vulnerabilities in complex codebases with an architecture it calls Agentic MapReduce: it maps signals across the repository, fans out bounded agents over shards, reduces the findings to a single report, and verifies serious vulnerabilities in isolated sandboxes before confirming them.
Cursor launched its iOS app in public beta, with management of cloud or local agents from the phone, updates through Live Activities, and pull request merging. And ZCode arrived on macOS, Windows and Linux, with GLM-5.2 tuned for the tool and a 1.5x usage quota for GLM Coding Plan subscribers.
Sources: Cognition, 30 June and 3 July — “Devin Fusion” and “Introducing Devin Security Swarm”; Cursor, 30 June — “Build from anywhere with Cursor for iOS”; report via TLDR AI, 2 July — “ZCode”.
6. Open source and efficiency: DSpark, LongCat-2.0 and Laguna XS 2.1
DeepSeek open-sourced DSpark, a framework that speeds up language-model inference by up to 85%. The system works like a scout that runs a few steps ahead, proposes the likely path, and lets the larger model quickly check which steps are safe, without changing what the model is trying to say.
Meituan launched LongCat-2.0 on its APIs, a 1.6-trillion-parameter Mixture-of-Experts model aimed at agentic coding, multi-step workflows and long-context processing. The model turned out to be the engine behind “Owl Alpha”, an unidentified model operating on OpenRouter that recently ranked among the top three by global daily volume.
Poolside published Laguna XS 2.1, a 33-billion-parameter Mixture-of-Experts optimised for agentic coding and long-horizon tasks, with a 5.4-point improvement on SWE-bench Multilingual to 63.1%. It ships under the OpenMDW-1.1 licence, with three quantized checkpoints, on Hugging Face and via API. And Sakana launched Fugu, with a 93.2 score on LiveCodeBench and pricing starting at $5 per million input tokens.
Sources: VentureBeat, 30 June — “DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%”; TestingCatalog, 1 July — “Meituan launches LongCat-2.0 1.6T parameter model on APIs”; Poolside, 3 July — “Introducing Laguna XS 2.1”; Implicator, 30 June — “Sakana Fugu launches with 93.2 LiveCodeBench score after Claude ban”.
7. Framework of the week: two data series on AI and employment
A joint study by Ramp and Revelio Labs analysed the relationship between firm-level generative AI investment and employment across more than 21,000 companies in the United States. Companies with high-intensity AI spending grew their overall headcount by 10.2% and entry-level positions by 12% over the two years following adoption. The study covers US companies.
For its part, the June 2026 report of the Anthropic Economic Index notes that AI computational costs correlate strongly with the economic value of tasks: higher-wage occupations consume up to 2.5 times more tokens.
As context, a VentureBeat article argued this week that coding agents have multiplied engineering output and shifted the bottleneck from writing code to deciding what to build, with more demand for profiles that combine technical fundamentals with product judgement.
Sources: Ramp and Revelio Labs, 2 July — “A new look at AI’s impact on jobs: firm-level AI spending and workforce adjustment”; Anthropic, 29 June — “Anthropic Economic Index June 2026 report”; VentureBeat, 29 June — “Claude Code turned every engineer into three. Now companies need more product thinkers”.
Closing
The six blocks cover the return of Claude Fable 5 and Mythos 5 after the export controls were lifted, the announcements of GPT-5.6 Preview, Grok 4.5 and Watermelon, Anthropic’s week of launches, the moves by Google and Meta in generative image and compute capacity, the run of coding agents from Cognition, Cursor and ZCode, and the open models from DeepSeek, Meituan, Poolside and Sakana. The framework of the week gathers the data published by Ramp and Revelio Labs and by the Anthropic Economic Index on AI and employment. The items reflect what the companies and the cited reports communicated; they do not include independent verification of the figures.