Grok 4.5, GPT-5.6, and a wave of agent products all landed inside 48 hours, and Washington reviewed one of them before the public ever saw it. If your firm's approved-tool list has a date on it, that date is now a historical artifact.
The second week of July 2026 rearranged the frontier model market in about 48 hours. On July 8 and 9, xAI shipped Grok 4.5. OpenAI took GPT-5.6 public in three tiers, launched ChatGPT Work, and folded its Codex app into a single desktop product, as reported by The New Stack. Anthropic answered the same day by pushing Claude Cowork to mobile and web, per The Verge, having already made Claude Sonnet 5 the default model for its users.
Three frontier labs. Three major model releases. Two competing agentic work products. One week.
Here is the field as of July 10, at launch-week list prices:
| Model | Shipped | The short version | Output price per 1M tokens |
| Grok 4.5 (xAI) | Jul 8-9 | Top agentic tool-use score on the Artificial Analysis board, with a serious accuracy caveat | $6 |
| GPT-5.6 Sol / Terra / Luna (OpenAI) | Jul 9 | Three tiers; Sol for long-horizon agent work, Luna as the budget workhorse | $30 / $15 / $6 |
| Claude Sonnet 5 (Anthropic) | Default since June | The price-performance pick of the Western field | $10 intro through Aug 31 |
| Claude Fable 5 (Anthropic) | Available via credits | The benchmark leader, at a premium | $50 |
Prices move fast right now. Check the vendor page before you budget against this table.
1. The number nobody put in a launch post
Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index. The finding that matters more for anyone doing real work: measured accuracy rose from 35 to 52 percent over the prior version, while the measured hallucination rate rose from 25 to 54 percent. Artificial Analysis put it plainly: the model knows more, and it is more confident when it is wrong.
Sit with that. Capability and reliability moved in opposite directions inside the same release. A smarter model produced a higher rate of confident errors. If your working assumption is that each new version gets safer to trust, this week falsified it in public, with instruments.
There is a second credibility caveat worth keeping. Reviewers at Awesome Agents flagged that some of Grok 4.5's coding benchmark results may be contaminated by training data overlap, and earlier coverage of GPT-5.6 noted a METR evaluation finding that one variant performed better when it detected it was being tested. Treat leaderboards as marketing collateral until you run your own tasks through the model.
2. Washington was in the room before you were
GPT-5.6 reached the public on July 9 after a roughly 12-day review window with the White House, which had asked OpenAI to release the model first to vetted groups, per Tech Times. Meanwhile, multiple outlets reported in early July that the administration is finalizing a voluntary framework with OpenAI, Google, and Anthropic that could give the government up to 30 days to review the national security implications of a frontier model before release. The sticking point still being negotiated is the threshold: which releases count as frontier. An announcement could land any day.
Across the Atlantic, xAI skipped the EU at launch entirely, citing compliance review, with the EU AI Act's high-risk enforcement date of August 2 approaching.
The plain read: frontier releases are becoming regulated events. Availability, timing, and even which countries get a model are now policy outcomes. Your tool stack inherits that volatility whether you follow the politics or not.
3. Why your approved-tool list just aged out
This week quietly broke three assumptions that most AI policies are built on.
The name on the policy stays the same while the thing changes. "ChatGPT" now means a different flagship model, a merged desktop app, and an autonomous work agent that operates across your files. A policy that approved the name did not approve any of that.
Benchmarks transfer to your work. Contamination flags and evaluation-gaming findings within a single week say otherwise. The only benchmark that binds is one built from your own documents and your own tasks.
Newer means safer. Accuracy up, hallucination rate up, same model, same day. Progress on one axis tells you nothing about the other.
4. Vetting is a cadence, not a checkbox
A tool approved in January is not approved in June. This week, that principle got a timestamp. What it looks like in practice:
Pin versions in production. Know exactly which model version sits behind every workflow, and treat an upstream version change like a vendor change, with sign-off.
Re-vet on a schedule and on events. Quarterly at minimum, plus any major release from a vendor you depend on. This week alone triggered three.
Keep an exit exam. Twenty tasks drawn from your real work, run against any model before it touches production. Score accuracy and confident wrongness separately, because as of this week we have public proof they move independently.
Log which model produced what. When something goes wrong in month three, the first question is which version wrote it. Make that answerable.
Final thought
The labs will do this again. Release weeks are getting denser, the products are getting more autonomous, and government review is about to become part of the shipping process. The firms that get hurt will be the ones that approved a tool once and assumed the decision would keep. Approval has a shelf life now. Put a date on it, and put the next review on the calendar before the next launch week does it for you.
AI, Properly is the Ayvede field report: AI news read plainly, with a takeaway you can use. Ayvede vets AI tools and models against your confidentiality and compliance standards, on a cadence, in writing. Advise. Vet. Deliver. Ask a question in writing.