This brief covers the trailing ~72 hours (July 3 – 6, 2026). Every item below was confirmed on the originating source’s own page, with a published date inside (or, where noted, just before) the window. With the US July 4th holiday weekend, the major labs’ newsrooms were quiet; the notable developments this cycle came from the practitioner community, plus a formal-verification model release from Mistral that landed just before the window opened.
Armin Ronacher documents a tool-calling regression in Anthropic’s newest models
Armin Ronacher · July 4, 2026
Flask creator Armin Ronacher published a detailed investigation showing that Anthropic’s newest models — Opus 4.8 and Sonnet 5, but none of their older siblings — sometimes call third-party edit tools with extra, invented fields that violate the tool’s JSON schema, even when the edit content itself is byte-correct. His hypothesis: post-training via reinforcement learning inside Claude Code (whose harness silently repairs malformed calls) gives the models a strong prior for Claude Code’s flat edit-tool shape, implicitly punishing alternative schemas used by other harnesses like Pi. Simon Willison amplified the finding the same day, asking whether third-party coding harnesses will need to ship model-specific edit tools.
“What surprised me is that this is getting worse with newer Anthropic models as both Opus 4.8 and Sonnet 5 show it but none of the older models. In other words, the SOTA models of the family are worse at this specific tool schema than their older siblings.” — Armin Ronacher
Source: Better Models: Worse Tools (see also Simon Willison’s commentary)
Simon Willison ships sqlite-utils 4.0rc2 “mostly written by Claude Fable” for an estimated $149.25
Simon Willison · July 5, 2026
Willison released sqlite-utils 4.0rc2 with the bulk of the work done by Claude Fable 5 via Claude Code, including a pre-release review that caught five release-blocker bugs — among them a data-loss bug where delete_where() never committed and poisoned the connection. He then had OpenAI’s GPT-5.5 review Fable’s work, which surfaced two further P1 transaction-handling issues that Fable confirmed and fixed. Using AgentsView he estimated the unsubsidized API cost of the session at $149.25, noting he upgraded to the $200/month Max plan ahead of July 7, when Fable’s subsidized subscription access ends.
“Over the course of 37 prompts, 34 commits and +1,321 -190 code changes over 30 separate files, we worked through the entire set of feedback in turn, making several other design improvements along the way.” — Simon Willison
Source: sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)
Still developing
Mistral releases Leanstral 1.5, an open Apache-2.0 theorem-proving model
Mistral AI · July 2, 2026
Just before this window opened, Mistral released Leanstral 1.5, a formal-verification model for Lean 4 with 119B total and 6B active parameters, free under Apache-2.0 with weights on Hugging Face and a free API endpoint. Mistral reports it saturates miniF2F at 100%, solves 587 of 672 PutnamBench problems at roughly $4 per problem, and sets new state-of-the-art results on the FATE-H (87%) and FATE-X (34%) abstract-algebra benchmarks. Beyond mathematics, an automated Rust-to-Lean verification pipeline built on the model flagged 47 violated properties across 57 open-source repositories, 11 of them pointing to genuine bugs — 5 previously unreported.
“Today, we are releasing Leanstral 1.5, a free Apache-2.0 licensed model with 119B total and only 6B active parameters, delivering a performance upgrade that makes formal verification more powerful and accessible than ever.” — Leanstral Team at Mistral AI
Source: Leanstral 1.5: Proof Abundance for All
This brief covers the trailing ~72 hours (July 3 – 6, 2026).
Primary sources: