Claude Haiku 5.5 and GPT‑6 in ChatGPT lead a busy 48 hours

Anthropic's new small model is built for high-volume work such as summarization, classification and subagents (helper agents a larger model delegates pieces of a task to). It has a 1M-token context window and is the first Haiku-class model to support the effort setting, which controls how much it reasons. Pricing for requests up to 100K tokens is $0.10 input and $0.50 output per million tokensVendor claim. Anthropic also halved cache-read pricing for Sonnet 5.5 on the same day.

Anthropic says it is about 75% cheaper than Haiku 4.5 on averageVendor claim. Moving the costliest steps of an agent pipeline (summarizing, compacting, lookups) onto this model could change your unit economics. It also reports 72.4% on the OSWorld 2.1 computer-use benchmarkVendor claim. The same day, a browser and computer-use toolset (beta) was added to the SDK, so this is a good time to try a low-cost browser automation agent.

GPT‑6 is now in the ChatGPT chat tab. Paid plans get GPT‑6 Sol, and Free and Go plans get GPT‑6 Luna. The main feature is Intelligent UI, where the model builds its answer out of interactive elements such as text, charts, buttons and forms (generative UI: a screen assembled on the fly for each question). Answers can also start streaming before the model has finished thinking.

Chat answers are turning into small apps. That could change the competition for product teams that ship lightweight tools like calculators or comparison tables. OpenAI says GPT‑6 Instant starts answering 44% sooner on questions that need web searchVendor claim. That figure is time to first response, not time to a complete answer.

Google is extending Gemini into a work agent that takes goals rather than step-by-step instructions. It hands work off to subagents, and its model picker includes third-party models such as Anthropic's Claude. Through MCP (Model Context Protocol, a standard for plugging tools into agents) servers it connects to Workspace, M365, Slack, Jira, Snowflake and more. Each agent gets its own Workspace account and email address, and its audit trail is recorded under the agent's name, not a person's.

The design treats the agent as a member of the organization rather than as a user's tool. That makes permissions, approval flows and audit logs the main question for anyone adopting it. Google also announced real-time spending limits and smart routing, which give enterprises a way to control multi-model costs.

Ultrafast mode, which runs GPT‑6.1 Sol up to 8x faster than standardVendor claim, is rolling out to the API, Codex and ChatGPT Work. It is called over WebSocket (an HTTP alternative exists). API pricing is $12 input and $60 output per million tokensVendor claim.

It is meant for work where waiting is the cost, like incident response, agents that operate apps directly, and real-time experiences. The per-token price is high, though. Routing only latency-sensitive steps to this mode makes more sense than sending every call through it.