AI Catchup

The AI Catchup -- July 10, 2026

By 4 min read

Welcome back. Two weeks ago we told you GPT-5.6 was a watch-this-space signal -- a limited preview for trusted partners, no prices, no numbers. The space just opened. OpenAI made the full family generally available, and this time it arrives with a price list, a benchmark table, and two platform changes that matter to anyone running production traffic. We lead with what changed since the preview, then follow a busy week on the Anthropic side: Cowork is escaping the desktop, API rate limits quietly went up, and Claude Code Artifacts reached individual plans.

Let us get into it.

This Week in AI

OpenAI made GPT-5.6 generally available (vendor post July 9) across ChatGPT, Codex, and the API, ending the limited preview that began June 26 at the request of the U.S. government. The family splits into three tiers, and for the first time we have real prices, per million tokens:

  • Sol, the frontier flagship OpenAI calls a step function better than GPT-5.5: $5 input, $30 output.
  • Terra, the balanced tier pitched as GPT-5.5-competitive at 2x lower cost: $2.50 input, $15 output.
  • Luna, the high-volume workhorse and OpenAI's most cost-efficient model: $1 input, $6 output.

The benchmark story centers on agentic work. Per OpenAI's published table, Sol scores 88.8% on Terminal-Bench 2.1 (against 88.0% for GPT-5.5), and a new "ultra" configuration that runs four agents in parallel pushes that to 91.9%. On OSWorld 2.0, OpenAI reports 62.6% while using 85% fewer output tokens than Claude Opus 4.8 -- read that as an efficiency pitch, not just a score.

Two platform changes ship alongside the models:

  • Prompt caching got predictable. GPT-5.6 supports explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate; cache reads keep the 90% discount. If you have been engineering around best-effort caching, you can now design for it deliberately.
  • Programmatic Tool Calling landed in the Responses API. Instead of round-tripping every tool call through the model, the model writes JavaScript that orchestrates your tools inside an isolated V8 runtime with no network access.

Also worth marking: OpenAI says Sol is launching on Cerebras at up to 750 tokens per second for select customers this month. The day-one move: test Claude Fable 5 orchestration with GPT-5.6 Sol workers -- you mitigate costs while maintaining strong agentic workflows, keeping the frontier planner on top and letting Sol run the long-horizon execution its Terminal-Bench numbers were built for.

The Bigger Move

Anthropic says Claude Cowork is coming to mobile and web (vendor post July 7): hand Claude a task at your desk, then pick up the finished work from your phone while Claude keeps going. It is a beta rollout over the next several weeks, starting with the Max plan and expanding from there. One caveat we verified in Anthropic's help docs: live artifacts remain desktop-only for now, so Cowork on web or mobile will not show the Live artifacts tab.

The direction of travel should look familiar. Last issue, Cursor put cloud agents on the iPhone; now Anthropic is moving Cowork the same way. The pattern across vendors is consistent: the agent runs somewhere else, and the phone is becoming the review surface.

Ship It This Week

Two direct wins you can act on today.

Give your voice agent reasoning at no extra cost. OpenAI added GPT-Realtime-2.1-mini to the API (vendor post July 6) with reasoning and tool use at the same cost as GPT-Realtime-mini. If you run a voice or streaming agent on the mini tier, the move is one line: swap the model string in staging and A/B the difference. You are adding planning and tool calls to your voice flow without touching the budget. One check first: the platform docs lagged the announcement, so confirm the exact model string and pricing before production.

Publish your first Claude Code artifact. Artifacts -- live web pages generated from a running Claude Code session -- now cover Pro and Max, not just Team and Enterprise. The whole flow is two steps: run /login in Claude Code (sessions authenticated with an API key cannot publish), then ask it to publish an artifact of your session -- a PR walkthrough, a dashboard, a release checklist. On individual plans the page stays private to you; Team and Enterprise keep org-wide sharing.

Quick Hits

  • Anthropic raised Claude API rate limits for all users. The developer announcement (vendor post July 2) raises limits across the board and simplifies usage tiers -- they are no longer based on API spend, and the newest Sonnet and Haiku models get 5x higher rate limits at the highest tier. Re-check your capacity assumptions in the Claude Console, and keep your 429 backoff logic anyway.
  • OpenAI open-sourced a slice of GeneBench-Pro. A research-level benchmark for whether agents can handle judgment-heavy computational biology -- messy data, revised assumptions, deciding when results are decision-ready. The full set is 129 questions; 10 representative case studies are on Hugging Face under the MIT License. Useful signal for anyone evaluating agents on real analysis work rather than one-shot Q&A.

That is it for this issue. Two weeks ago GPT-5.6 was a preview for trusted partners; today it is three models with public prices and a benchmark table. If a teammate is still budgeting API spend on GPT-5.5 assumptions, forward this -- Terra changes the math.

Until next week, stay caught up.

Get the weekly AI Catchup

Tools, practices, and what matters -- in your inbox every Monday.