AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, marking new versions of its AI models with significant improvements across coding, knowledge work, and scientific research. Fable 5.1 costs approximately 25% less than Fable 5 for typical workloads due to 75% reduction in cache read pricing, with savings up to 45% for highly agentic tasks. The models show performance gains across multiple benchmarks: Terminal-Bench-Science 0.1 (52.6%), Terminal-Bench 4.0 (60.9% for Mythos 5.1), Humanity's Last Exam (65.0% with tools), and CursorBench 3.2.0 (73.4%), outperforming predecessors like Fable 5 and Opus 5. Mythos 5.1 demonstrated advanced scientific capabilities including protein design with hit rates near 50% across 12 targets (versus typical 10-15%), computational speedups up to 2.5x for deep learning models, and creation of high-resolution Venus elevation maps. The models include improved safeguards with 60% fewer false positives in cybersecurity and 85% fewer interventions for benign biology queries, alongside new Enterprise Frontier Safeguards for data privacy and anti-distillation mechanisms.

GPT-6 Astra set new records on Epoch AI's Capabilities Index and multiple benchmarks including math and continual learning tasks following pre-release evaluation in early September 2026. Epoch AI launched FrontierMath Erdős, featuring 68 unsolved Erdős problems where AI systems write solutions in Lean; GPT-6 Astra solved 2 of 68 problems (3%), marking the first model to solve any of them. The benchmarking hub tracks 390 models across 80 benchmarks including math, coding, and knowledge tasks, with evaluation methodology using the Inspect framework and publicly accessible evaluation code and model outputs for transparency.

TypeSafe AI released Jev on September 15, 2026, the first "System One" model that returns typed decisions with calibrated probabilities instead of text generation. The model costs $0.042 per million input tokens with free output, delivers responses in 70–500ms, and guarantees schema conformance. On TypeSafe's workflow evals, Jev achieves 76.0% accuracy while costing 25–145x less than frontier models and running 193.6x faster, using a parallel sampling architecture and reinforcement learning for calibrated decisions (RLCD) rather than autoregressive generation. The model is available via REST API and official Python and JavaScript SDKs, with no hallucination possible since outputs are enumerated in advance; TypeSafe acknowledges pricing may be subsidized and expects it to decrease over time.

LiteLLM released benchmarks for its Rust AI Gateway, showing 0.66 ms p99 latency overhead—3.5× lower than competing gateways like Portkey (2.29 ms) and Bifrost (4.54 ms)—with throughput of 2,800+ requests per second at ~21% CPU utilization, measured using the open standard AI Gateway Bench. The gateway supports 140+ LLM providers and 1,892 unique models, with day-zero support for new model releases and is available as open-source (MIT-licensed) for self-hosting or enterprise deployment with SSO and audit logs.