OpenAI has launched GPT-6 Astra, aimed at computer-use work: filling forms, updating CRM records, calendar tasks, browser QA and longer agent loops. It is rolling out to ChatGPT Plus, Pro, Business and Enterprise over the coming days, and to the API as gpt-6-astra, with AWS/Bedrock as well. On Enterprise, Astra starts off. An admin has to enable it. That is a useful default if you are not ready to give an agent write access yet.

For most EU SMB and Spanish agency desks, the useful comparison is cost against Claude Fable 5.1, not the launch rhetoric.

Sticker price. Astra’s API Standard rate is $10 per million input tokens and $50 per million output. That matches Fable 5.1’s list price. It is also a clear increase from GPT-5.6 Sol’s promotional $4/$20 class. Same headline numbers as Anthropic; different behaviour once you run real agent sessions.

Cache (this is where long loops diverge). Fable 5.1 cut cache reads by 75% to $0.25 per million tokens. Anthropic estimates about 25% lower cost on typical token-billed workloads, and up to around 45% on highly agentic, cache-heavy runs. Astra still lists separate cache rates; launch pricing tables put cached input around $1 per million and cache writes around $12.50. If your agent keeps re-reading the same HubSpot, Ads or docs context, Fable’s cache line is much cheaper. Astra’s Fast mode offers up to 2× speed at 2× Standard price. Fine for a timed job; expensive if you leave it on by habit.

Do not treat either vendor’s “cheaper agent” claim as your invoice. Artificial Analysis found Fable 5.1 can still cost more per task than Fable 5 at max effort, because it used roughly 1.7× the output tokens even after the cache cut. Astra will have the same issue on long computer-use runs that dump screenshots, tool traces and CRM writes. The honest approach is to replay a week of your own loops on both models and compare.

Data retention. OpenAI offers zero data retention for eligible API customers. Anthropic’s Enterprise Frontier Safeguards (customer-cloud storage with a ZDR-like posture) are due later this fall, with interim ZDR for eligible customers on Fable. If Spanish clients will not put CRM exports on default US retention, settle eligibility before any agent gets production data.

Deployer checklist (secondary to price, still required). If Astra can update customer records or talk to shoppers, you are in GDPR deployer territory: purpose limitation, access scope, logging, and Article 50 questions for AI-assisted customer text. Keep Enterprise access off until you have (a) a cost sheet and (b) a narrow permission set. Prefer read-only connectors first.

What to do this week: estimate Astra Standard vs Fast vs Fable 5.1 using your real cache hit rate and output volume; leave Astra disabled on Enterprise until that estimate exists; if you still connect HubSpot or similar, scope writes tightly and log them. Same sticker price. Different bill once agents run for an hour.