AI developer tools · What shipped

Claude Fable 5 leads Cursor's CursorBench v3.2 coding benchmark at 70.5%, followed by GPT-5.6 Sol (67.2%) and Grok 4.5 (66.7%). The benchmark evaluates 12 models on ambiguous, multi-file agentic coding tasks from real Cursor sessions, with results updated July 10, 2026. CursorBench is currently displayed for reference but excluded from BenchLM's overall scoring formula due to being a first-party benchmark, though the top performers remain closely clustered within 3.8 points in the leadership tier.

Claude Fable 5, developed by Anthropic, achieved a new high score of 161 on the Epoch Capabilities Index (ECI), surpassing GPT-5.5 Pro by 1 point and marking Anthropic's first lead on the ECI in over a year. Epoch AI recently expanded its benchmarking hub by tracking 13 new evaluations, with 7 incorporated into the ECI, covering areas including agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics across updates from June 15-July 1, 2026.

v1.49.0 of Gentle AI released with performance improvements focused on bounded delegation and code review workflows. Key changes include a streamlined review lifecycle (scope → implement → test → one review → freeze → one fix → targeted verify), native OpenCode delegation for read-only repository exploration, optional Pi CodeGraph lifecycle management, and durable pre-push review authority that maintains approval validity after commits. The release was validated across Ubuntu, Arch, and Fedora with all lifecycle requirements and agent parity tests passing. v1.46.0 introduced a rewritten 4R review framework replacing recall-at-any-cost loops with deterministic routing and structural budgets. The update implements precision-gated reviews where trivial diffs skip review entirely, standard diffs run one lens, and high-risk diffs run the full risk assessment set. Findings now require adversarial verification before driving code changes, with severity floors allowing only BLOCKER and CRITICAL findings into fix loops and convergence budgets capping remediation at two fix rounds. Runtime parity was propagated across Claude, Cursor, Kimi, OpenCode, Gemini, and other supported adapters.

Claude Opus 4.7 and Opus 4.8 lead autonomous Rails coding benchmarks at 97 and 95 scores respectively, with Opus 4.7 maintaining the top position at ~$1.10/run. GPT 5.4 xHigh and 5.5 xHigh achieve tied 97 and 96 scores but at 15× and 10× the cost. Among cost-efficient tier-A options, Kimi K2.6 reaches 87/100 at ~$0.30/run with correct RubyLLM API implementation. The benchmark reveals that structural completeness (file counts, test counts, artifact checklists) does not predict runtime correctness—most models hallucinate the RubyLLM API, with only Claude Opus, Claude Sonnet, and GLM 5 using the correct chat API pattern. DeepSeek V4 Pro reaches tier-A performance (84-89/100) when run through Claude Code via the deepclaude shim at $3.14/run, fixing earlier opencode incompatibility. Local llama-swap models show significant improvements over Ollama for reliability: Qwen 3 Coder Next completes at 17m with 1675 files, Qwen 3.5 models both complete successfully. The two-phase validation approach (code generation plus Docker/local boot verification) materially improves run quality by forcing actual deployment validation rather than stopping after code generation.

Meta released Muse Spark 1.1, a coding model priced at $1.25/$4.25 per 1M input/output tokens, undercutting Claude Opus 4.8 by 75% on input costs and 83% on output costs. The model shows competitive performance on agentic coding benchmarks like MCP Atlas and JobBench, with particular strength in tasks requiring planning and orchestration across external applications and services, including zero-shot generalization to new tools and custom skills.