Frontier models work, but supervision tools lag
Claude Opus 4.6, GPT-5.3-Codex and OpenAI Frontier, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
Anthropic
Claude Opus 4.6
DEPLOYif you run long agentic coding tasks
WHAT SHIPPEDClaude Opus 4.6 is available today on claude.ai, the API, and all major cloud platforms at $5/$25 per million tokens. It adds a 1M token context window in beta, leads Terminal-Bench 2.0 and Humanity's Last Exam, and beats GPT-5.2 by around 144 Elo points on GDPval-AA.
THE CATCHThe 1M context window is beta, and Opus 4.6 can add cost and latency on simpler tasks because it thinks more deeply by default. Anthropic recommends dialing effort down from high to medium when the model overthinks.
IF YOU RUNa team doing multi-step coding, code review, or long-context retrieval, this model's 76% on the 8-needle 1M MRCR v2 versus Sonnet 4.5's 18.5% is the number that matters
SOURCEAnthropic, Introducing Claude Opus 4.6
OpenAI
GPT-5.3-Codex
DEPLOYif you pay for ChatGPT and use Codex
WHAT SHIPPEDGPT-5.3-Codex is available with paid ChatGPT plans everywhere Codex runs: the app, CLI, IDE extension, and web. It sets a new industry high on SWE-Bench Pro and Terminal-Bench, runs 25% faster, and matches GPT-5.2 on GDPval.
THE CATCHOpenAI classifies GPT-5.3-Codex as High capability for cybersecurity tasks and routes some requests it flags as elevated cyber risk to GPT-5.2 automatically. Security researchers can apply for full access through Trusted Access for Cyber.
IF YOU RUNa development team that wants one model for coding, debugging, deploying, writing PRDs, and analyzing data, this is the first Codex model trained to identify software vulnerabilities
SOURCEOpenAI, Introducing GPT-5.3-Codex
OpenAI
OpenAI Frontier
WATCHif you run enterprise agents across many systems
WHAT SHIPPEDFrontier is available today to a limited set of customers, with broader availability coming over the next few months. HP, Intuit, Oracle, State Farm, Thermo Fisher, and Uber are among the first adopters.
THE CATCHFrontier is not generally available yet. The page says 75% of enterprise workers say AI helped them do tasks they couldn't do before, but the platform itself is limited to a small customer set for now.
IF YOU RUNa large enterprise with agents scattered across clouds, data platforms, and applications, Frontier's shared context, identity, permissions, and boundaries are the pitch, but you cannot buy it today
SOURCEOpenAI, Introducing OpenAI Frontier
The Deploy Log
Models and APIs
| DEPLOYif you build voice agents or transcription Voxtral Transcribe 2. Voxtral Mini Transcribe V2 is available now via API at $0.003 per minute, and Voxtral Realtime is available via API at $0.006 per minute and as open weights on Hugging Face under Apache 2.0. Mistral |
| DEPLOYif you build Apple platform apps Xcode Claude Agent SDK. Xcode 26.3 introduces a native integration with the Claude Agent SDK, giving developers subagents, background tasks, plugins, and visual verification with Xcode Previews inside the IDE. Anthropic |
| WATCHif you localize UI elements on 4K screens Holo2-235B-A22B Preview. Holo2-235B-A22B Preview achieves 78.5% on Screenspot-Pro in agent mode within 3 steps and 79.0% on OSWorld G, and is available on Hugging Face. Hugging Face |
Tools and products
| DEPLOYif you ship production apps from prompts New v0. The new v0 can import any GitHub repo, connect to Snowflake and AWS databases, and let anyone on a team open PRs against main and deploy on merge. Vercel |
| WATCHif you generate synthetic training data SyGra Studio. SyGra 2.0.0 introduces Studio, an interactive environment that turns synthetic data generation into a visual canvas with guided model forms, data previews, and live execution streaming. Hugging Face |
| DEPLOYif you run Python on Vercel Python 3.13 and 3.14. Builds and Functions now support Python 3.13 and Python 3.14 alongside the previously supported Python 3.12, with the default switching to Python 3.14 in the coming months. Vercel |
| DEPLOYif you run builds on Vercel Pro Turbo build machines default. Turbo build machines are now the default for all new Pro projects and projects upgrading from Hobby to the Pro plan, delivering 30vCPUs and 60GB of memory. Vercel |
| DEPLOYif you use TRAE for coding TRAE integration. ByteDance's coding agent TRAE now integrates both AI Gateway and direct Vercel deployments, bringing unified AI access and instant production shipping to over 1.6 million monthly active developers. Vercel |
Left out
- Snowflake partnership: the publisher's page does not show it available yet
- Koa zero-configuration support: it did not carry a number that moved
- Asana search in Notion: it did not carry a number that moved
- Parallel on Vercel Agent Marketplace: its own name for the thing was too long to print
- Parallel web search on AI Gateway: its own name for the thing was too long to print
- Sanity on Vercel Marketplace: it did not carry a number that moved
- Claude Opus 4.6 on AI Gateway: its own name for the thing was too long to print
- Codex app: it did not carry a number that moved
The pattern
This week the frontier models stopped being the bottleneck. Claude Opus 4.6 and GPT-5.3-Codex both ship with agentic coding, long-context, and knowledge-work numbers that would have been a full release cycle apart a year ago, and both are available today at unchanged or faster pricing.
The bottleneck has moved to supervision and integration. OpenAI's Frontier is limited to a small customer set, the Codex app and v0 are rebuilding interfaces for managing parallel agents, and Vercel's marketplace items are all about wiring models into existing workflows. The models can do the work; the tools for directing them are still catching up.