GPT-5.6 now rivals Claude Fable 5 for a third of the cost

OpenAI has released GPT-5.6, a new family of models that now matches Anthropic's Claude Fable 5 on most work while costing roughly a third as much to run.

That sounds like a reason to move your firm from one lab to the other. For most firms it is not. The calmer response is to work out which model suits which kind of task, and to test that on your own work before you change anything.

What GPT-5.6 is, and what changed

GPT-5.6 comes in three versions. Luna is the cheapest and fastest, Terra is the everyday model with quality close to the previous generation, and Sol is the most capable. OpenAI released the family for general use on 9 July, rolling it out globally over the following day.

The number that matters for a leader is not any single benchmark. It is the combination of two things. On the independent Artificial Analysis intelligence index, the top model Sol came within a point of Claude Fable 5 at about a third of the cost per task.

Near-top quality did not get much better this week. It got a lot cheaper.

What it costs next to Fable 5

On the rate card, GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output, against Fable 5 at $10 and $50. That is half the price on input, and lower on output too.

The rate card is the wrong thing to compare, though. Models now spend very different amounts of hidden reasoning on the same request, so a per-token price tells you less than it used to, as one developer who has been using both models pointed out.

What you actually pay is the cost of a finished task. On the same independent test, Sol came in near $1 a task where Fable was closer to $2.75, and the cheaper Terra and Luna cost less again.

One more thing changed on price. Fable 5 has moved from being bundled into paid Claude plans to metered, pay-per-use billing, so every Fable task now carries a cost as well. The practical lesson is the one in our note on whether an AI tool pays for itself. Judge it by what a real piece of your work costs to finish, not by the headline rate.

Can your firm actually use it?

For ordinary work, both models are available to you today. GPT-5.6 is in ChatGPT, in OpenAI's coding tool and in the developer API, and Fable 5 is on paid Claude plans. Nothing about the launch stops a professional-services firm using either one this week.

There is one caveat, and it is about the very top of the range rather than daily use. Sol is the most cyber-capable model OpenAI has shipped. Its most sensitive settings went through a United States government review before release, and they now sit behind extra identity checks, with access restricted in some places.

That is a control on the most advanced hacking capability, decided under US rules. It does not oblige a firm here to do anything, and it does not touch the drafting, analysis and research a leader actually reaches for.

Where each model is stronger

The honest picture is that these two are close, and both are a clear step above the next rank of models. Neither is simply better than the other.

Fable 5 still leads where the work is long, self-contained and demanding. It tops the same independent index overall, and it scores higher on realistic knowledge-work quality. It is also well ahead on repository-level software engineering, where it solved 80 per cent of a hard coding benchmark to Sol's 65.

The developer quoted above, who uses Anthropic's model daily, said Sol had not beaten it on his complex coding.

GPT-5.6 leads elsewhere. It is stronger on fast agent-style tasks that call tools and browse, it produces more polished slides and documents, and it does its work with fewer tokens, in less time, at lower cost. That efficiency, rather than a jump in raw intelligence, is the story of this release.

Which to reach for, by task

A simple rule covers most firms. Reach for Fable 5 on the long, judgement-heavy jobs you can brief fully and hand over, then check hard. Think of the due-diligence memo that ends in a recommendation, the section of an audit that has to be right, or a serious piece of software work.

Reach for GPT-5.6 on the faster, back-and-forth work, on tasks that call tools and browse, and on building a deck. Where cost or volume matters, the cheaper Terra and Luna handle everyday drafting for far less.

Picture a mid-sized accountancy firm at month end. The partner drafting a client's board commentary, thinking aloud and revising as the numbers land, may prefer the quicker, step-by-step feel of GPT-5.6. The same firm running a full technical review of a lease portfolio, where one missed clause is the whole risk, may want Fable 5 on the long task. A senior person should check the output either way.

Hold that rule loosely. The gap between the two is small on the everyday writing and summarising that fills most days, and the standard tiers already do that well. There is little sense in rebuilding a workflow around a one-point difference on a leaderboard.

Why this is not a moment to switch everything

A benchmark measures the model. It does not measure whether the model suits your documents, your house standard and the people who will use it. That fit is what decides whether a tool earns its place, and the only way to see it is to run real work through both and read the results as a client would.

Switching has costs that a price comparison hides. People have to relearn a tool. Prompts and workflows have to be rebuilt, and the firm takes on the risk of depending on one supplier. Fable 5 was withdrawn worldwide for about three weeks in June under an export ruling, a plain reminder to keep a fallback, as we set out in our note on vendor continuity.

The prices will keep moving too. Anthropic will probably respond, and this week's gap may not be next month's. A decision made purely on today's cost is one you will be making again soon.

The throughline is simple. The model drafts, and a person still decides. Running this kind of comparison on a leader's own work, which model for which task and at what real cost, is what our 1-to-1 AI lessons for leaders are built around. It is the same structured trial we described for testing a top model on real client work.

Questions leaders ask

Is GPT-5.6 better than Claude Fable 5?

Not outright. On the main independent index Fable 5 is about a point ahead, while GPT-5.6 Sol sits just behind at roughly a third of the cost and leads on agent-style coding and efficiency. They are strong at different things, so what helps is knowing which fits a given task, rather than which wins overall.

How much does GPT-5.6 cost compared with Fable 5?

The top tier lists at $5 per million input tokens and $30 per million output, against Fable 5 at $10 and $50. The fairer measure is the cost of a finished task, where independent testing put the top GPT-5.6 model near $1 against Fable's $2.75, with the smaller tiers cheaper again. Confirm it on your own work before you budget.

Should a firm switch from Claude to ChatGPT?

Not on this news alone. Run a few real tasks through both, keep whichever fits each kind of job, and avoid leaning so hard on one supplier that an outage or a price change leaves you stuck.

Is the most capable version available to everyone?

All three tiers are available now in ChatGPT and the developer API. Only the most cyber-capable settings of the top model are gated, behind identity checks brought in after a US government review, and that gating does not affect ordinary business use.

The next step is small. Pick two or three real tasks this week, run them through both models, and compare the finished work and what each one cost you to get there. If you would like help turning that into a clear rule for which model your firm uses, and where, book a discovery call and we will work through it with you.