Anthropic Releases Claude Opus 5.5 at Lower Prices
Anthropic released Claude Opus 5.5 (claude-opus-5-5) on September 22 as the first model in its 5.5 family. List pricing is $4 input and $20 output per million tokens, 20% below Opus 5. Cache reads fall 60% to $0.20 per million tokens. Anthropic says typical workloads cost about 40% less to run than on Opus 5, and that the model generates output roughly 30% faster. A fast mode is available at $8/$40 per million tokens for up to 2.5x speed. The model has a 1M-token context window.
On benchmarks, Anthropic reports 66.4% on Terminal-Bench 4.0, up from 52.3% for Opus 5, and says Opus 5.5 matches Claude Fable 5.1 on most work at a lower price. Customer anecdotes in the launch post include a 680,000-line code migration completed in under a day. On the safety side, Anthropic reports improved resistance to prompt injection and says the model is 85% less likely than Opus 5 to try to get around containment boundaries. External pre-deployment evaluations came from METR and others. Cybersecurity-heavy requests are routed to an older model unless the user is enrolled in Anthropic's Cyber Verification Program.
The model is available on the Claude Platform, AWS, Google Cloud, and Microsoft Azure. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow within weeks. Claude Code 2.1.280 made Opus 5.5 its default Opus model on launch day.
Read more — Anthropic
OpenAI Launches GPT-6 Sol and Luna at Half the Price, Codex CLI 0.156–0.157 Follow
OpenAI released GPT-6 Sol and GPT-6 Luna on the same day, September 22. The two models replace the GPT-5.6 pair at roughly half the price. Sol, aimed at complex coding and professional work, costs $2 input and $10 output per million tokens, down from $4/$20. Luna, for fast high-volume work, costs $0.10/$0.50. Both have a 1.1M-token context window and are exposed in the API as gpt-6-sol and gpt-6-luna. OpenAI also says prompt caching now hits more often by default, with up to a 90% discount on cached input and explicit cache breakpoints for agent loops.
OpenAI's benchmark claims focus on cost per task. Sol at extra-high effort scores 33.2% on AutomationBench at $0.27 per task. OpenAI puts Claude Opus 5 at maximum effort on 26.9% at 11 times that cost. On DeepSWE, Sol scores 68.8% against Claude Fable 5's 69.9% at about 80% lower cost per task. These are vendor-reported comparisons against the previous generation of Claude models, not Opus 5.5.
The Codex CLI picked up the models quickly. Version 0.156.0 (September 22) added an optional fullscreen UI via /tui with transcript search and mouse selection. It also turned on voice conversations by default with an F8 toggle, added a /usage dashboard for token totals, and enabled worktree support by default. Version 0.157.0 (September 25) added GPT-6 Sol and Luna with Amazon Bedrock support and migration prompts for older models. It also made fullscreen transcripts the default, added an f shortcut to fork a conversation while keeping drafts, and starts a background server automatically for eligible interactive sessions.
Read more — MarkTechPost
Claude Code 2.1.280–2.1.283: Model Allow/Deny Lists, Prompt Audits, and MCP Telemetry
Claude Code shipped four releases this week. Most of the changes are for teams administering the tool at scale. Version 2.1.283 (September 25) adds availableModelsMatch and deniedModels managed settings, so organisations can pin or block specific model versions. It adds a /doctor prompt-audit command that reviews CLAUDE.md files and prompting patterns. MCP tool, WebFetch, and WebSearch outputs are now emitted as OpenTelemetry span events, and gateway requests carry an x-claude-code-prompt-id header so all requests from one prompt can be grouped.
Version 2.1.281 adds assume_role and Bedrock guardrail configuration for Bedrock upstreams on the Claude apps gateway. It also adds MCP URL-mode elicitation on connections using the 2026-07-28 protocol revision, and MCP server validation in claude plugin validate. Version 2.1.282 changes the default model on Pro and Team Standard plans from Sonnet to Opus. It also adds a maxProseWidth setting for wide terminals and improves startup time by deferring UI loading.
Each release also fixes session-resume bugs. Resumed sessions no longer re-send earlier turns in a changed form. Very large sessions resume fully instead of restoring only the last few messages. A dropped MCP server no longer invalidates the prompt cache mid-conversation. Stdio MCP servers are now shut down when a session ends.
Read more — Claude Code Docs
Hugging Face Transformers Can Now Load and Serve llama.cpp GGUF Quants
Hugging Face announced on September 22 that Transformers can run llama.cpp GGUF quantized checkpoints directly. Loading uses the familiar API: AutoModelForCausalLM.from_pretrained(repo, gguf_file="...Q4_K_M.gguf"). transformers serve "model_id:file.gguf" exposes the model through an OpenAI-compatible endpoint. The post benchmarks Q4_K_M, Q5_K_M, and Q6_K variants of Qwen3.5-4B (2.74–3.53 GB) and reports throughput close to llama.cpp itself.
The initial release has limits. It currently requires Apple Silicon, a build of Transformers from main until the next tagged release, and one of the two latest PyTorch versions. Support is optimised first for the Qwen3.5 architecture.
This matters because GGUF is the most common format for locally distributed models, but until now researchers had to leave the Python and PyTorch stack to use it. Running the same quantized file in Transformers makes it possible to evaluate a quantized model with existing pipelines, attach hooks or modify layers, and compare quantization levels without converting weights.
Read more — Hugging Face
Links & Sources
You May Also Enjoy
