Fable 5 ships, vanishes under the first export control on model weights, and returns. Plus Sonnet 5, Nemotron 3, and the stateless MCP SDKs.
In June the frontier grew gatekeepers: governments and vendors inserted access controls between model and builder, while the agent stack hardened into boring, governable infrastructure with MCP betas, IDE integrations, and management controls. Every link goes to the primary source.
A new tier above Opus: 1M context, 128k output tokens, $10/$50 per million tokens, adaptive thinking always on. Integration-relevant: Fable 5 ships safety classifiers that return stop_reason "refusal" as an ordinary response, plus a server-side fallbacks parameter; Mythos 5 is the same model without classifiers, limited-release via Project Glasswing.
Approaches Opus 4.8 capability at $2/$10 per million tokens, becomes the default on Free and Pro, and is available via the API, AWS, and Microsoft Foundry. Ships with cyber safeguards on by default; the workhorse price point for agent fleets.
Led by Nemotron 3 Ultra, a 550B mixture-of-experts model with 55B active as a Mamba-Transformer hybrid with 1M context, built for long-running tool-using agents under the permissive OpenMDW license. Regarded at release as the strongest open-weights American model and a serious option for self-hosted agent backends.
Python v2, TypeScript v2, Go, and C# implement the protocol without the initialize handshake and without sessions, so remote servers scale behind plain load balancers. Breaking changes included: Python's FastMCP becomes MCPServer and the TypeScript SDK splits into server and client packages. June was the feedback window before the July 28 final.
Codex passed 5 million weekly users and gained plug-ins for analytics, creative work, sales, design, and finance, plus Sites as hosted interactive apps. The signal: Codex is becoming a general-purpose agent harness, not just a coding tool.
Copilot Chat in JetBrains can drive the Claude Code CLI from its agent picker; the same changelog adds org-level shared custom agents and message queueing in the Copilot CLI. Cross-vendor agent interop, landing inside a mainstream IDE.
Auto mode reaches Bedrock, Google Cloud Agent Platform, and Microsoft Foundry; acceptEdits now always prompts before writing executable-adjacent files like .zshenv or .npmrc, and orgs can enforce an approved version range. Later in the month, Sonnet 5 became the default model.
Instead of compacting agent context at fixed intervals, the model decides when to summarize, for example after a subtask: up to 18 points better on math and 5 to 9 on search, at 30 to 70 percent lower token cost. Directly applicable if you manage long-horizon agent context.
Measures frontier models and agent harnesses on granular research-lifecycle tasks; the best configuration reached only 68.3 percent and failed on details obvious to experienced researchers. Useful calibration for any claim of "autonomous research".
A voluntary framework for developers to give the government up to 30 days of pre-release access to covered frontier models, plus a Treasury-coordinated vulnerability clearinghouse. No mandatory licensing, but it shapes when new frontier models reach the API you build on.
The first export-control directive aimed at model weights as such: Anthropic would have needed a license before any transfer to foreign persons, found selective blocking infeasible, and disabled both models for everyone. If you depend on a single frontier model, this month taught you to build fallback paths.
OpenAI announced GPT-5.6 Sol, Terra, and Luna but complied with a US government request to roll out first to a small partner group; the public release followed in July. June API roadmaps had to plan around staged availability.
After the export-control suspension, Fable 5 returned globally on July 1, alongside a proposal developed with Amazon, Microsoft, and Google to score jailbreak severity consistently across the industry. Worth reading for how refusal behavior at the frontier is likely to be standardized.
