Anthropic Launches Faster, Cheaper Sonnet 5.5 as Vespper Debuts DOCX MCP

Anthropic says Sonnet 5.5 keeps the same per-token price as Sonnet 5—$2 per million input tokens, $10 per million output tokens and $0.20 per million cache reads—but typically uses fewer tokens and runs more than 30% faster in output speed, making it the fastest Sonnet model to date and about 30% cheaper for the same work. The model is available Monday on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Opus 5.5 costs twice as much per token, and Anthropic says Haiku 5.5, its cheapest model, is planned for the coming weeks.

The company positions Sonnet 5.5 for everyday tasks with clear scope, such as fixing coding bugs, creating polished documents, slides and spreadsheets, and refining user interfaces. Anthropic said early testers described it as a better collaborator than Sonnet 5, with clearer writing and speed suited to quick iteration on less complex tasks. Theo Chu, a research product manager at Anthropic, told CNBC that Sonnet is aimed at cost-conscious customers who may not need as much intelligence as Opus, and that it is suited to routine tasks that need execution but not Opus-level judgment.

On benchmarks, Anthropic said Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, an agentic coding evaluation, compared with 10.3% for the previous model. It scored two points below Opus 5.5 on GDPval-AA, a test of real-world occupational work, and became the first Sonnet model to beat Pokémon Red using only screenshots, a test of long-horizon work and image understanding. TechCrunch reported that Anthropic's benchmarks show Sonnet 5.5 performing better than Opus 5.5 on agentic coding, likely because it can spawn multiple agents without exceeding cost limits.

Anthropic said Sonnet 5.5 has cybersecurity capabilities comparable to Opus 5, making it the first Sonnet model to launch with cyber safeguards and fallbacks similar to those used for its most capable models, including Fable and Mythos-class models. Those safeguards may trigger on prompts involving cybersecurity, biology or other sensitive content, but Anthropic said most software development and life sciences work will be unaffected. Biology safeguards remain unchanged from Sonnet 5. The company also said Sonnet 5.5 does not advance the frontier of its model capabilities, so most alignment testing focused on a targeted set of risks that apply to models at any capability level.

SiliconANGLE reported that Sonnet 5.5 includes invisible text watermarking to comply with global regulations, including the European Union's AI Act. Anthropic has said the watermarking does not affect text quality or readability but increases the likelihood that AI-generated text can be detected. Zendesk Director of AI Abhinay Kathuria said in a statement that Zendesk fed Sonnet 5.5 hundreds of real support use cases and that it made fewer wrong decisions than the Claude models in production, processing tickets 20% faster.

In a separate launch, Vespper, a Y Combinator F24 company, introduced an MCP that lets AI agents edit Word documents. According to a Hacker News post by founders Dudu and Topaz, the tool is powered by a fine-tuned model and is currently 3× faster, 2× cheaper and more accurate than the closest alternative. The founders said they spent a year building an AI document editor for pharmaceutical companies, where users preferred working with their own Word templates. They described Word documents as zip files of verbose OOXML XML files, where even small changes require complex operations, causing agents to burn time and tokens.

Vespper's approach gives agents HTML instead of raw OOXML. The agent makes find-and-replace edits, and Vespper reconciles those edits back into the original .docx file. The company chose HTML over Markdown because it is structurally closer to OOXML and because CSS associates styles with elements roughly as OOXML does. Vespper wrote its own DOCX-to-HTML converter; the conversion is lossy, but the HTML is only a projection for the agent, while the original file remains the source of truth and is mutated in place. The MCP exposes three tools—read, search and edit—and reconciliation is handled by a fine-tuned 3–8B base model with a LoRA adapter. Vespper said its internal benchmark shows the tool is more accurate than the DOCX skill and raw python-docx while being about 2× cheaper and 3× faster, taking three tool calls per task at the median versus 10 for the DOCX skill and 13 for Office CLI. The founders said images and comments are not supported yet, and they are seeing use by legal tech companies powering live-editing flows in Office.js.