AI developer tools · What shipped

Moonshot AI released Kimi K3 on July 16, 2026, a 2.8-trillion-parameter open-source model with a 1M-token context window and native multimodal capabilities. Built on two new architectural components—Kimi Delta Attention (KDA) and Attention Residuals (AttnRes)—K3 achieves roughly 2.5x the scaling efficiency of its predecessor. Full model weights will be released by July 27, 2026. On long-horizon coding benchmarks, K3 leads all tested models: Terminal-Bench 2.1 (76.1%), DeepSWE (67.3%), and Program Bench (79.0%). It trails on PostTrain Bench and overall UX compared to Claude Fable 5 and GPT-5.6 Sol, but competes closely on SWE Marathon and achieves 87.1% on browse agent tasks. The model uses a Stable LatentMoE framework activating 16 of 896 experts per forward pass and includes case studies demonstrating superior GPU kernel optimization and the ability to design functioning chips autonomously. Kimi K3 is available via OpenAI-compatible API at $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output, with reported cache hit rates above 90% on coding workloads. The API supports up to 1,048,576 max completion tokens, configurable reasoning effort (low/high/max), and automatic context caching for prefixes over 256 tokens. K3 requires thinking history preservation in agent harnesses and exhibits sensitivity to history truncation; it is compatible with Kimi Code and verified compatible with Claude Code when configured as a backend model.

Meta launched Muse Spark 1.1 on July 9, 2026, its first closed-weight frontier model with a paid developer API priced at $1.25 per million input tokens and $4.25 per million output tokens. The 1M-token multimodal model features parallel sub-agents, active context compaction, Computer Use capabilities across desktop/browser/mobile, and drop-in SDK compatibility with OpenAI and Anthropic libraries. Meta claims state-of-the-art performance on MCP Atlas, JobBench, Humanity's Last Exam, and FinanceBench, though a Terminal-Bench 2.1 compute cap dispute (6 CPU cores, 8GB RAM) significantly asterisks the headline benchmark claims. The agentic orchestration architecture is the genuine competitive angle; coding performance trails Opus 4.8 and GPT-5.5. This marks Meta's strategic shift from the "open-weights company" (Llama) to a hosted, metered API model targeting enterprise developers alongside OpenAI and Anthropic. The API is currently US-only in developer preview. Meta scored D+ on the FLI Safety Index and is excluded from the White House voluntary 30-day pre-release review framework—the lowest safety governance score among major Western labs. The positioning targets teams seeking agentic capability below frontier-tier pricing, with SDK compatibility enabling cheap A/B testing against competitors.

Anthropic released Claude Opus 5 on July 24, a mid-tier model that closes the capability gap with their frontier Fable 5 while maintaining Opus 4.8 pricing ($5/M input, $25/M output tokens). Opus 5 scores 61 on the Artificial Analysis Intelligence Index, achieves state-of-the-art results on Frontier-Bench v0.1 and GDPval-AA, and notably scores 30.2% on ARC-AGI-3—roughly three times higher than competing models. Key technical improvements include mid-conversation tool swapping without invalidating prompt cache, automatic fallback to Opus 4.8 on safety classifier refusals (85% less restrictive than Fable 5), and no 30-day data retention requirement, making it practical for regulated industries. The release establishes a clear three-tier production pricing ladder: Sonnet 5 for throughput, Opus 5 for everyday agentic work at half Fable's cost, and Fable 5 for frontier reasoning tasks.