AI developer tools · What shipped

Artificial Analysis Intelligence Index v4.3.2 has been released, incorporating 10 evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The benchmark tracks model performance across weighted cost per task, output speed, and capability indexes covering finance, strategy, legal, healthcare, engineering, and economics domains. Gemini 3.8 Flash is highlighted as positioned in the most attractive cost-performance quadrant.

OpenAI released GPT-6 Sol and GPT-6 Luna, two new models in the GPT-6 family positioned below the flagship Astra. Both models are live in the OpenAI API today. Sol and Luna API prices have been cut by 50% compared to GPT-5.6 pricing: Sol input/output costs $2/$10 per 1M tokens (down from $4/$20), while Luna costs $0.10/$0.50 per 1M tokens (down from $0.20/$1.20). On AutomationBench 1.0.6, Sol at xhigh effort scores 33.2% at $0.27 per task, outperforming Claude Opus 5 at 11.1x lower cost. On DeepSWE v1.1 coding benchmarks, Sol scores 68.8% at roughly 80% lower cost than Claude Fable 5, while Luna scores 66.6% comparably to Opus 5 and Fable 5 at medium effort but costs 93-96% less per task. On OSWorld 2.0 offline, Sol matches Opus 5's performance at 80% lower cost. OpenAI improved prompt caching with up to 90% discounts on cached input reads and new developer controls including a caching dashboard, diagnostics tools, and explicit breakpoints.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, with Fable 5.1 priced 25% lower than Fable 5 for typical workloads due to 75% reduction in cache read pricing ($0.25 per million tokens), and up to 45% savings for highly agentic work. Input pricing remains $10 per million tokens and output $50 per million tokens. Fable 5.1 demonstrates improved performance across multiple benchmarks including Terminal-Bench-Science 0.1 (52.6%), Terminal-Bench 4.0 (55.8%), CursorBench 3.2.0 (73.4%), and Humanity's Last Exam (60.9% with tools), outperforming Fable 5 across coding, knowledge work, and agentic tasks. The models ship with enhanced developer capabilities including improved safeguards that reduce false positives by 60% in cybersecurity contexts, support for vulnerability discovery (defensive work only), and new Enterprise Frontier Safeguards allowing zero-data-retention deployments on customer-controlled cloud infrastructure. Mythos 5.1 adds permissive safeguards for vetted cybersecurity and life sciences professionals through trusted access programs developed in partnership with the US government.

Ember-1 is a new specialized model from Fireworks Research delivering Kimi K3's quality with 40% fewer tokens by learning to cut unnecessary reasoning while preserving essential thinking. The model was developed through 50+ training experiments across mathematics, coding, instruction following, conversation, search, tool use, and software engineering tasks, using Fireworks Serverless Training to move from research to launch faster. On external benchmarks including Doximity's Bedside Bench (500 clinical cases across 10 categories), Ember-1 set a new Pareto frontier for cost-per-task performance against GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5, achieving 35–50% token reduction compared to K3 at maximum reasoning effort across seven industry benchmarks including Terminal Bench, SWE-bench Verified, and DeepSWE 1.1. Live A/B tests with two production customers on coding workloads confirmed approximately 35% fewer tokens per task at comparable quality, with one customer now running Ember-1 in live production. Ember-1 is available today as a Research Preview release on Fireworks Serverless with a two-week access model, with training support also launching to enable enterprises to build customized token-efficient models.