AI Radar for Business — Thursday, July 9, 2026

We launch the daily Radar on the most competitive day of the year for AI models: OpenAI opens GPT-5.6 to the public today with official pricing, SpaceXAI beat them to the punch by hours with Grok 4.5 — an "Opus-class" model at $2 per million tokens that EU companies cannot use yet — and Anthropic is pulling its most powerful model out of Claude subscriptions and moving it to pay-per-use, with a reprieve until July 12 after user pushback. The takeaway if you run automation: cost per token is the new battleground, and assigning a model to each task — with a budget — is no longer optional.

GPT-5.6 goes public today: Terra at half the cost, Luna from $1 per million tokens

After the US government review that kept it limited to select partners, OpenAI is rolling out all three GPT-5.6 tiers today with official API pricing and a progressive global release: Sol, the top of the range, at $5 per million input tokens and $30 for output; Terra, the everyday workhorse, at $2.50 and $15 — GPT-5.5-level performance at half the price, by OpenAI's own account; and Luna, the budget tier, at $1 and $6. For your business: it's time to re-price your workflows — whatever runs on GPT-5.5 today should cost roughly half on Terra, and high-volume tasks (classification, extraction, replies) can drop to Luna. Assign tiers per task before the invoice decides for you. Source

Grok 4.5: an "Opus-class" model at $2 per million — but the EU has to wait

SpaceXAI — formerly xAI, now under the SpaceX umbrella — has launched Grok 4.5, trained in partnership with Cursor and aimed at coding, agents and knowledge work, at $2 per million input tokens and $6 for output. The company positions it against Anthropic's Opus models, claiming (on its own numbers) around four times fewer output tokens on coding tasks. It is already available on Grok Build, across all Cursor plans and through its API console — but EU access is not live yet, with mid-July as the expected date. For your business: more competition at the high end pushes prices down across every provider — use it to renegotiate your cost per task. But if you operate in Europe, don't build anything critical on Grok yet: a provider that ships late to the EU with no firm date is a continuity risk, not a bargain. Source

Fable 5 leaves Claude subscriptions: grace period until July 12, then pay-per-use

Anthropic was set to pull Fable 5 from Claude's paid plans in the early hours of July 8, but extended included access until July 12 after user complaints; from then on the model is billed through usage credits at API prices — $10 per million input tokens and $50 for output — while Sonnet 5, Opus 4.8 and Haiku 4.5 remain included. A company engineer would only say it returns to subscription plans "when capacity allows," with no date attached. For your business: you have until Sunday to measure which tasks genuinely need the top model and which can fall back to Sonnet 5 with no noticeable loss. And note the structural lesson: an AI provider's "all-inclusive" plan no longer guarantees access to its best model — treat model access as a variable cost in your budgets and contracts. Source

What to watch tomorrow

The first independent benchmarks for GPT-5.6 and Grok 4.5 — and whether Luna punches above its price; Google's response, with Gemini 3.5 Pro still undated; and, in Europe, the AI Act Omnibus reaching the Official Journal. This is the first issue of the daily series: from now on, the Radar lands every morning with the AI news that changes something for your business.