OpenAI ships a safer default while Anthropic buys compute
GPT-5.5 Instant, GPT-Realtime-2 and Claude usage limits, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
OpenAI
GPT-5.5 Instant
DEPLOYif you use ChatGPT daily
WHAT SHIPPEDGPT-5.5 Instant replaces GPT-5.3 Instant as the default model for all ChatGPT users and in the API as chat-latest. It produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering medicine, law, and finance, and reduced inaccurate claims by 37.3% on conversations users flagged for factual errors.
THE CATCHThe model recovers from its own initial mistake in a math example, catching that x=3 fails when plugged back into the original equation, identifying the algebra error, and solving the corrected quadratic. GPT-5.3 Instant stopped early and incorrectly concluded there was no real solution.
IF YOU RUNa team that relies on ChatGPT for everyday answers, the default model is now more accurate and less verbose without any action from you
SOURCEOpenAI, GPT-5.5 Instant: smarter, clearer, and more personalized
OpenAI
GPT-Realtime-2
DEPLOYif you build voice agents
WHAT SHIPPEDGPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper are available in the Realtime API. GPT-Realtime-2 scores 15.2% higher on Big Bench Audio than GPT-Realtime-1.5 and 13.8% higher on Audio MultiChallenge for instruction following. Context window increases from 32K to 128K.
THE CATCHGPT-Realtime-Translate supports more than 70 input languages and 13 output languages, priced at $0.034 per minute. GPT-Realtime-Whisper transcribes speech live as the speaker talks at $0.017 per minute.
IF YOU RUNcustomer support or live events, you can now add live translation and streaming transcription without separate vendors
SOURCEOpenAI, Advancing voice intelligence with new models in the API
Anthropic
Claude usage limits
DEPLOYif you hit Claude rate limits
WHAT SHIPPEDAnthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, removed the peak hours limit reduction on Claude Code for Pro and Max accounts, and raised API rate limits considerably for Claude Opus models, all effective May 6, 2026.
THE CATCHThe SpaceX deal gives Anthropic access to more than 300 megawatts of new capacity, over 220,000 NVIDIA GPUs, within the month. This joins an up to 5 gigawatt agreement with Amazon and a 5 GW agreement with Google and Broadcom.
IF YOU RUNClaude Code on a Pro or Max plan, your peak hours limit reduction is gone and your five-hour rate limit is doubled
SOURCEAnthropic, Higher usage limits for Claude and a compute deal with SpaceX
The Deploy Log
Models and APIs
| WATCHif you run red teaming GPT-5.5-Cyber. GPT-5.5-Cyber is rolling out in limited preview to defenders responsible for securing critical infrastructure, with more permissive behavior for authorized red teaming, penetration testing, and controlled validation, while GPT-5.5 with Trusted Access for Cyber remains the recommended starting point for most security workflows. OpenAI |
Tools and products
| DEPLOYif your workspace uses Custom Agents Custom Agent Directory. Notion added a Custom Agent Directory in your Library where you can browse all your workspace's agents, pin favorites, and create new ones to automate your team's busywork 24/7. Notion |
| DEPLOYif you run security scans deepsec. Vercel open sourced deepsec, a security harness powered by coding agents that runs on your own infrastructure, uses Opus 4.7 at max effort and GPT 5.5 at xhigh reasoning, and scales up to 1,000+ concurrent sandboxes on Vercel's codebases. Vercel |
| DEPLOYif you need cross-platform history Chat SDK conversation history. Vercel's Chat SDK now supports cross-platform conversation history through transcripts and identity options, letting the same user keep up to 200 messages per user for 30 days across every platform adapter. Vercel |
Left out
- Finance agent templates: the publisher's page does not show it available yet
- Trusted Contact: the publisher's page does not show it available yet
- Chat SDK Messenger adapter: the publisher's page does not show it available yet
- Chat SDK web adapter: the publisher's page does not show it available yet
- Vercel metrics in CLI: it did not carry a number that moved
- Production-only credentials: it did not carry a number that moved
- Git committer auto approval: it did not carry a number that moved
- Vercel Flags JSON values: it did not carry a number that moved
- Custom Agent admin controls: it did not carry a number that moved
- Plan Mode: it did not carry a number that moved
- ChatGPT ads: it did not carry a number that moved
The pattern
OpenAI shipped three defaults in one issue: GPT-5.5 Instant as chat-latest, GPT-Realtime-2 in the Realtime API, and GPT-5.5-Cyber in limited preview. Each one removes a decision a team previously had to make. GPT-5.5 Instant cuts hallucinated claims by 52.5% on high-stakes prompts and recovers from its own algebra error where GPT-5.3 Instant stopped early. GPT-Realtime-2 raises instruction following by 13.8% on Audio MultiChallenge and expands context from 32K to 128K, with translation at $0.034 per minute and transcription at $0.017 per minute. GPT-5.5-Cyber opens more permissive behavior for authorized red teaming while GPT-5.5 with Trusted Access for Cyber stays the recommended starting point. Anthropic counters not with a model but with capacity: Claude Code rate limits double on May 6, 2026, peak hours limits disappear for Pro and Max, and the SpaceX deal adds over 220,000 NVIDIA GPUs and more than 300 megawatts within the month, joining 5 GW agreements with Amazon and Google and Broadcom.
The surrounding ecosystem turns those defaults into persistent infrastructure. Notion's Custom Agent Directory lets a workspace browse, pin, and create agents that automate busywork 24/7. Vercel open sourced deepsec, which runs Opus 4.7 at max effort and GPT 5.5 at xhigh reasoning across 1,000+ concurrent sandboxes on Vercel's own codebases. Vercel's Chat SDK keeps up to 200 messages per user for 30 days across every platform adapter. Together these say the next wave is not a single model release but a set of safe defaults plus compute and memory that make those defaults stick.