GPT-5.6 Sol, the sun that throws shade at Fable

On Thursday, July 9, OpenAI opened GPT-5.6 to the public after two weeks of restricted access for a handful of hand-picked partners^[1]. I don't cover every model release, there's one a week and this blog would not survive the pace, but I do follow Anthropic and OpenAI more closely than the rest, since each still leads its side of the field. This time the calendar deserves a look of its own, because on July 12, three days later, Anthropic is due to pull Fable 5, its best model, out of its subscription plans. OpenAI is aiming at the king, and it's aiming at the wallet.

Sol, Terra, and Luna, OpenAI discovers astronomy

The GPT-5.6 family comes in three models named after celestial bodies, and the company promises the names now mark durable product lines rather than disposable versions^[1:1]. Sol, the flagship, targets advanced reasoning, software development, scientific research, and complex agents. Terra aims at everyday work with GPT-5.5 performance at half the cost, while Luna competes on speed and rock-bottom pricing. For a company that spent years naming its models like motherboard part numbers, the sudden turn to poetry is a welcome development.

Money is where it gets interesting. On the API, a million input tokens^[2] on Sol runs $5 and a million output $30, with Terra at half those rates and Luna at a fifth^[3]. More to the point, Sol is available in ChatGPT starting with the $20 Plus plan, with the Sol Pro variant reserved for the pricier tiers^[4], and all three models landed in GitHub Copilot the same day^[5]. Keep those numbers in mind, they make the rest of this story considerably funnier.

One detail of the launch deserves a pause, because it continues a story told on this very blog. The late-June restricted preview happened at the US government's request, a June executive order now requires labs to submit their most powerful models thirty days before release, and the Commerce Department tested GPT-5.6 before clearing the wide rollout^[3:1]. OpenAI clearly took notes during Anthropic's misadventure, when Washington pulled the plug by decree in June, and the government-in-the-loop arrangement I wrote about last week took barely a week to become standard procedure. Launching a model now looks a lot like launching a rocket, government safety review included, which is fitting given the names.

The benchmarks are a tie, for whatever that's worth

On paper, the duel with Fable 5 ends in a draw. Artificial Analysis, the only neutral evaluator to have measured both models, scores Fable at 60 and Sol at 59 on its composite index, a gap that lives comfortably inside the margin of error^[6]. Beyond that, each side waves the numbers that flatter it. Sol claims 88.8% on Terminal-Bench 2.1 against Fable's 83.4%, while Fable rules SWE-Bench Pro at 80.3% on a test where OpenAI has prudently declined to publish a score^[6:1]. Two vendors, two measuring tables, and almost no shared benchmark^[7], which makes the comparison about as rigorous as a fishing contest where everyone brings their own scale.

Which is a good moment to repeat a conviction of mine, the numbers don't settle it. A few years back, when ChatGPT was widely considered far ahead of Claude, I never warmed to OpenAI's model, which hallucinated more and followed instructions the way a teenager follows curfew. Benchmarks measure test performance, not the feeling of working with a tool that understands what you asked for, and that feeling, for me, has always sat on Claude's side. A statistical tie on the charts can hide a clear winner in daily use, in either direction.

Meanwhile, on planet Anthropic

The real battlefield is elsewhere, in how each company charges for access to its best model. Anthropic plans to remove Fable 5 from its subscriptions and gate it behind prepaid credits at $10 per million input tokens and $50 per million output, double Sol's price. The deadline, originally July 7, has already slipped to July 12, and the company swears the retreat is a temporary capacity measure, with the model returning to subscriptions once the hardware catches up^[8]. I believe the intention. I just don't see what world the plan survives in. Pulling your best model from subscriptions at the exact moment your competitor slides its own into a $20 plan hands them a ready-made sales pitch, and once you've extended the deadline once, it gets awkward not to extend it forever. Competition, it turns out, has its perks, mostly for us.

The economics tilt the same way. OpenAI's tokens cost half as much and the model reportedly burns fewer of them per task^[6:2], while on Anthropic's side you need a Claude Max plan at roughly a hundred euros a month, minimum, to avoid hitting the usage limits within the hour, especially if you vibecode^[9] a little too enthusiastically, as I do. I love these models and I keep coming back to them, but the bill does raise my eyebrow now and then.

So I'm keeping my Anthropic subscription, out of attachment as much as habit, but OpenAI just landed a punch that puts it squarely back in the race. I should make time for a little infidelity and test Sol with my own hands, on my own projects, my fussy instructions, and my marathon coding sessions, instead of everyone else's benchmarks. If the experiment teaches me anything, you'll read about it here. Sol may be throwing shade at the fable, but it remains to be seen which of the two ends up lighting my desk.