Reviewed and written up.
Added Oct 8, 2026
Claude Haiku 5.5 is Anthropic's October 7, 2026 small model, replacing Haiku 4.5 as the recommended fast, cheap Claude. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output, 90% below Haiku 4.5 and the same as OpenAI's GPT-6 Luna; above 100,000 tokens the price is $0.50 and $2.50. Cache reads are $0.01 per million. Anthropic says about 90% of Haiku requests fall under the 100,000-token line, so the average cost falls by around 75%. In Anthropic's tests it scores 1620 on GDPval-AA v2.1 against Luna's 1437, 72.4% on OSWorld 2.1 against 48.9%, and 39.2% on Terminal-Bench 4.0 against 16.4%. It is the first Haiku with adjustable effort levels, from low to max, and supports computer and browser use in beta through the SDKs. Output is text only.
Why: It resets what a small model costs and what it can do at that price: on Anthropic's tables it beats GPT-6 Luna on every shared benchmark at the same per-token rate, and the independent Intelligence Index puts its max-effort setting above Luna's too. For high-volume classification, extraction, routing and sub-agent work, it is the new default to test first.
Added Oct 7, 2026
Mistral Large 4, nicknamed Le Chonk, is Mistral's flagship. The public preview opened on October 6, 2026. On October 29 Mistral published the weights of a further-trained checkpoint. The model card describes a natively multimodal mixture-of-experts model with 1.05 trillion total parameters, 52 billion active, a 1.6 billion parameter vision encoder, and a 1 million token context, trained on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. Mistral's weights post says this checkpoint scores 41% on Terminal-Bench 4.0, up 13 points from the preview, and 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than GLM-5.3, which it calls a tie. The same post sets the API price at $0.70 per million input tokens, $0.07 per million cached input tokens, and...
Why: The weights are out, so a European frontier model can be downloaded and run without Mistral serving it. The licence is a modified MIT: use, changes, and commercial products are allowed, and a business that offers the model as a service needs a separate commercial licence. Coding and cyber figures here are Mistral's, and the DeepSWE number is labelled as the preview.
Added Oct 7, 2026
Beam is the first model from Reflection AI, the Nvidia-backed New York lab founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Announced October 5, 2026, it is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, a 1-million-token context, and text-only input and output. Reflection reports 44.4% on DeepSWE v1.1, 80.1% on Terminal Bench v2.1, 80.9% on SWE-bench Verified and 90.5% on GPQA Diamond, broadly level with Z.ai's GLM-5.2 and behind Qwen3.8 Max and Kimi K3, and claims it reaches GLM-5.2's scores with 3 to 4 times less inference compute. The weights went up on Hugging Face on October 16 under Apache 2.0, in BF16, FP8 and NVFP4, with day-0 support in Transformers, vLLM and SGLang, a Rust-based agent harness, and...
Why: The US has had few serious open-weight models since Llama stalled, and Beam is a credible one under a fully permissive licence. It does not lead its Chinese peers, and Reflection's own tables say so. It is listed for the efficiency claim and the Apache 2.0 licence, ranked below the open models it trails until it is independently tested.