. Updated on . 6 min read.

GPT-5.6 Sol throws shade at Fable

GPT-5.6 Sol costs half of Fable 5 and comes with the $20 plan, three days before Anthropic pulls its model from subscriptions. The timing is nobody's accident.

I don't cover every model release. There's one a week, and this blog wouldn't survive. But I do follow Anthropic and OpenAI more closely, since they're each at the top of their game, and this time the calendar got involved. On Thursday, July 9, OpenAI opened GPT-5.6, its new family of models, to the public after two weeks reserved for a few hand-picked partners1. On July 12, three days later, Anthropic is supposed to pull Fable 5, its best model, out of its subscriptions. OpenAI is going after the king, and it's going for the wallet.

The Sun, the Earth, and the Moon lined up against a starry sky, labeled Sol, Terra, and Luna.
The solar system according to OpenAI (image OpenAI).

Sol, Terra, and Luna, OpenAI discovers astronomy

The GPT-5.6 family has three models named after celestial bodies, and the company promises the names now mark lasting product lines rather than throwaway versions1. Sol, the flagship, is for heavy reasoning, software development, research, and complex agents. Terra promises GPT-5.5 performance at half the price, for everyday work. Luna goes for speed and rock-bottom pricing. For a company that spent years naming its models like motherboard part numbers, the sudden poetry is nice to see.

On price, Sol costs $5 per million tokens2 in and $30 out, Terra half that, and Luna a fifth3. Most importantly, Sol comes with ChatGPT's $20-a-month Plus plan4. All three models also showed up the same day in GitHub Copilot, Microsoft's coding assistant5.

The launch picks up a story I've been telling here. The restricted preview in late June 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 signing off3. OpenAI clearly learned from Anthropic's misadventure, when the government pulled the plug in June. And the government in the loop I wrote about last week is already standard procedure. Launching a model now looks like launching a rocket, safety review included, which works out given the names.

The benchmarks are a tie, so what

On paper, the showdown with Fable 5 ends in a draw. The only neutral evaluator to test both, Artificial Analysis, gives Fable 60 and Sol 59 on its overall index, a gap well inside the margin of error6. Beyond that, each side waves the numbers that suit it. Sol claims a win on a terminal-commands test, while Fable rules a test of fixing real bugs where OpenAI has wisely declined to publish its score6. Two vendors, two scales, and almost no shared benchmark7. It's about as rigorous as a fishing contest where everybody brings their own scale.

Which is a good excuse to repeat something I believe, numbers aren't everything. A few years ago, when ChatGPT was considered way ahead of Claude, I never clicked with OpenAI's AI. It hallucinated more and followed instructions with the attitude of a teenager. Benchmarks measure tests, not the feeling of working with a tool that gets what you're asking. And for me, that feeling has always leaned toward Claude. A tie on the charts can easily hide a real winner in daily use, either way.

Meanwhile, on planet Anthropic

The real fight is over how to charge for the best model. Anthropic planned to pull Fable 5 from its subscriptions and only offer it through prepaid credits, at twice Sol's price. The deadline, first set for July 7, already slipped to the 12th. The company swears it's temporary, until the hardware catches up8. I'll take them at their word on the intent. But I don't see what world this plan works in. Pulling your best model from subscriptions right when your competitor slips theirs into a $20 plan is handing them a sales pitch on a silver platter. And once you've extended once, it gets awkward not to keep extending forever. Competition has its upside, mostly for us.

The money points the same way. OpenAI's tokens cost half as much, and the model reportedly uses fewer per task6. With Anthropic, you need a Claude Max plan at around a hundred euros a month at minimum to avoid hitting the limits within an hour, especially if you vibecode9 a bit, like I do. I love these models and I always come back to them, but the bill sometimes makes me raise an eyebrow.

So I'm keeping my Anthropic subscription, out of attachment as much as habit. But OpenAI just landed a punch that puts it back in the race. I should allow myself a little cheating and test Sol myself, on my projects, my fussy instructions, and my endless coding sessions, instead of relying on everyone's benchmarks. If I learn anything, you'll read about it here.

Update, July 12. Anthropic pushed the deadline back a second time, and Fable 5 now stays in subscriptions until July 1910. I wrote yesterday that once you've extended a deadline, it gets awkward not to extend it again. I didn't expect to be proven right this fast.

Notes

  1. OpenAI, "GPT-5.6", July 9, 2026, and "Previewing GPT-5.6 Sol", June 26, 2026. ↩ ↩2

  2. The token is the unit language models count in, a chunk worth about three quarters of a word on average. Labs bill per million tokens read (input) and produced (output). Fable 5 on prepaid credits costs $10 in and $50 out. ↩

  3. Engadget, "OpenAI gets permission to roll out GPT-5.6 to the public on July 9", July 2026, and CNBC, "OpenAI to publicly release GPT-5.6", July 8, 2026. ↩ ↩2

  4. 9to5Mac, "OpenAI unveils ChatGPT Work agent, GPT-5.6 models now available", July 9, 2026. The Sol Pro variant stays reserved for the pricier plans. ↩

  5. GitHub, "OpenAI's GPT-5.6 Sol, Terra and Luna are now available in GitHub Copilot", July 9, 2026. ↩

  6. Drawpie, "GPT-5.6 Sol vs Claude Fable 5: Cost, Benchmarks and What the Tests Actually Show", July 2026. Sol posts 88.8% on Terminal-Bench 2.1 against Fable's 83.4%, and Fable 80.3% on SWE-Bench Pro. ↩ ↩2 ↩3

  7. Benchmarks are standardized test batteries that grade models on code, reasoning, or general knowledge. Each lab mostly publishes the ones it shines on, like a résumé that quietly forgets the internship that went badly. ↩

  8. The New Stack, "Anthropic gives Claude subscribers five more days with Fable 5", July 2026. ↩

  9. Vibecoding is programming by describing what you want to an AI that writes the code, moving forward by conversation rather than by hand. Great for productivity, brutal on the token meter. ↩

  10. The Fable 5 promotional access now runs through July 19, 2026, on Pro, Max, Team, and Enterprise plans, as Anthropic's support page details. ↩

Read next…

  1. Opus 5.5 and GPT-6 Sol cut the price per token but not the bill

  2. Mistral raises three billion and Anthropic raises 65

  3. Nvidia buys Hugging Face and France had nothing to sell