AI developer tools · What shipped

Claude Science, an AI workbench for scientists, launched and is now available with customizable features integrating tools and packages researchers commonly use, along with auditable artifacts and flexible access to computing resources. Separately, Fable 5 returns globally on July 1, and Anthropic is proposing an industry-wide framework for scoring jailbreak severity in collaboration with Amazon, Microsoft, Google, and other partners.

Claude Fable 5 achieved a new high score of 161 on the Epoch Capabilities Index (ECI), surpassing GPT-5.5 Pro by 1 point and marking the first time Anthropic has led the benchmark in over a year. Epoch AI recently began tracking 13 new evaluations and added nine external benchmarks spanning agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics, with 7 of the new evals incorporated into the ECI.

little-coder is a coding agent tuned for small local models, achieving significant performance gains through architectural adaptation. The project recently published benchmark results across multiple frameworks: on Aider Polyglot, a 9.7B Qwen model achieved 45.56% (v0.0.2), with the larger Qwen3.6-35B-A3B variant reaching 78.67% (v0.0.5). On Terminal-Bench 2.0, Qwen3.6-35B-A3B scored 24.6% ± 3.2 (rank 120) and the smaller Qwen3.5-9B variant scored 9.2% ± 2.4 (rank 142), both accepted to the official leaderboard. On GAIA's validation set, Qwen3.6-35B-A3B achieved 40.0% (66/165 tasks) with per-level breakdowns of L1 60.4%, L2 37.2%, L3 7.7%. The v0.1.0+ release migrates from a Python substrate to pi, an agent framework providing the core loop, multi-provider support, TUI, and extension architecture. All benchmarks ran on consumer hardware (i9-14900HX, 32 GB RAM, 8 GB VRAM RTX 5070) without cloud inference. The detailed research methodology is documented in the accompanying Substack article "Honey, I Shrunk the Coding Agent."

JetBrains is rolling out AI for Teams and Organizations starting July 2026, introducing vendor-agnostic capabilities including team automations and cloud agents for long-running tasks, JetBrains Context to provide repository intelligence for more efficient code understanding, and JetBrains Central for organization-wide governance and cost management across multiple AI tools. The company is transitioning from AI licenses to flexible on-demand AI credits valid for twelve months, allowing organizations to manage AI adoption at scale while developers continue using their preferred tools through open integrations via MCP and ACP protocols.

As of July 8, 2026, GPT-5 leads the MATH-500 benchmark for competition mathematics problems with a 99.4% score, followed by o3 and Grok 3 Mini both at 99.2%. The benchmark evaluates 111 models on multi-step reasoning across algebra, geometry, number theory, and calculus. Other top performers include Claude Sonnet 4 Thinking at 99.1%, Grok 4 at 99.0%, and o4 Mini at 98.9%, with pricing data available to compare performance against cost.