On September 22–23, Anthropic and OpenAI rolled out new flagship model generations almost simultaneously — Claude Opus 5.5 on one side, and both GPT-6 Sol and GPT-6 Luna on the other. On paper, it's another round in the quality race: Anthropic points to a sharp jump in performance (in one test, the model migrated a 680,000-line codebase in under a day — work that normally takes weeks), while OpenAI emphasizes that its new models "lead across the cost-intelligence curve." But the quality headline isn't the real story here. The price is.
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. GPT-6 Sol comes in at $2 and $10 respectively. And GPT-6 Luna, OpenAI's smaller model, costs just $0.10 and $0.50 per million tokens. For comparison: a year or two ago, top-tier models at this level cost 5-10 times more on input tokens. This is happening against a rather awkward backdrop: just weeks before the release, Anthropic CEO Dario Amodei was publicly calling for a slowdown in AI capability growth, so the industry could catch up on risk safeguards. Both companies then released more powerful, cheaper models at exactly the moment when, by their own account, they should have been easing off.
Why a cheaper flagship model changes the economics of small-business automation
Until now, a small business owner often faced a binary choice: use a cheap, not always reliable model for routine tasks — sorting emails, drafting replies — or pay a premium for a top-tier model only where accuracy really matters: legal wording, complex calculations, work with large amounts of context. A 5-10x price drop on the flagship model erases that dilemma. It's now possible to keep a "top-tier" model in production for tasks where the cost of a mistake is high, without treating every API call as its own line item.
A practical example: a real estate agency with 6-8 staff runs a bot that parses incoming leads from the website and messaging apps, identifies the client's budget and preferred area, cross-checks it against the property database, and drafts a reply for the agent with three matching listings. At the old pricing for models of this class, a monthly bill for 3,000-4,000 inquiries could reach €150-200 — and the agency owner would deliberately simplify the bot's logic just to avoid paying for "excess intelligence." At the new pricing for Opus 5.5 or GPT-6 Sol, the same volume of inquiries costs €25-40, freeing up budget to add finer personalization — for instance, factoring in six months of a client's conversation history instead of just their last message.
What's worth checking before moving workflows onto the new models
There's a less rosy side to this, too. First, the published benchmark results are vendor showcases, not independent audits: migrating 680,000 lines of code in a day sounds impressive, but it says nothing about how the model handles your specific domain — say, standard contracts under Portuguese law or industry-specific terminology. Second, the contradiction itself — calls to "slow down" alongside the simultaneous release of more powerful models — is a signal that the industry still hasn't worked out clear safety rules for agentic systems that get access to real customer data and payment tools. Businesses connecting these models to a CRM or billing system should build in human review of critical actions rather than trusting the model outright, however impressive the benchmark. A reasonable minimum is a two-to-three-week test period on real but non-critical requests, with manual spot-checks of at least a tenth of the responses, before letting the model send emails or edit customer records without human confirmation.
There's a third point worth flagging — the sheer speed at which model generations turn over. If Opus 5.5 and GPT-6 Sol become outdated as fast as their predecessors did (and the refresh cycle has shrunk to roughly six months over the past year), businesses will need to budget not for a one-off rollout, but for a regular review of the bot's architecture: which tasks can move to a cheaper model, and which still need top-tier capability.
At Dayava, we connect exactly this kind of model to business workflows — from handling incoming leads to internal analytics — and we recalculate the economics for each client's actual volumes, not average figures from press releases. If a similar solution interests you for your business, leave a request at dayava.pt/contactos/.
Source: Hypertext, "After calls to slow down, Anthropic and OpenAI release new models"