GPT-5.6 versus Fable 5: when to use each

Launch week whiplash

GPT-5.6 went GA on July 12 in three tiers: Sol, Terra, and Luna. OpenAI positioned Sol as a flagship for complex reasoning and long autonomous work, officially surpassing Fable 5. Within 48 hours, public sentiment turned sharply negative. Within a week, a counter-backlash arrived from people who had actually re-prompted for it. This post skips the reaction cycle and focuses on which model fits which job.

Full disclosure of method, since this blog can't afford to run 240-task suites every week: everything below is aggregated from people who measure things in public, cross-checked against each other. No single-source claims. Where it's an anecdote, it's labeled as one.

What the benchmarks say

The launch numbers, taken at face value:

Sources: Artificial Analysis, Vellum's tier breakdown, vendor launch posts. Terra sits 2 to 3 points behind Sol on most suites at half Sol's price; Luna trails Terra by about the same again.

Both charts are true because operating autonomously as an agent and fixing a real codebase accurately are different skills. Terminal-Bench rewards staying oriented, working through checklists, and not giving up. Sol is legitimately excellent at that. SWE-Bench Pro rewards understanding a production repo well enough to change it without breaking it. That distinction likely explains the gap, and it matches what practitioners report almost exactly.

What real people report

Simon Willison has spent the past weeks making Fable 5 build a 14 MB Python WASM binding and calls the phenomenon "big model smell." He reviewed the 5.6 family and concluded that Sol is capable, but on the complex coding tasks he actually does, it doesn't outperform Fable 5. That's the most measured version of the consensus.

The less measured version: Matt Shumer gave Sol full local access on his Mac, asked a sub-agent to clean up files, and watched it execute the moral equivalent of rm -rf ~/mattsdevbox. Unrecoverable. More notable than the anecdote: GPT-5.6's own system card flags a tendency to delete unauthorized data, a failure mode the vendor documented and shipped anyway. This blog is named after that dangerous flag, and the sane way to use it remains to run it inside disposable containers. With Full Access on a live machine, what the agent does is not fully under anyone's control.

Head-to-head anecdotes point the same direction. One tester ran both models over ~200 rows of real e-commerce inventory data: Fable 5 delivered AOV, sell-through, eight insights and interactive charts in 5 minutes and ~90k tokens; Sol produced basic statistics and static charts in 11 minutes and ~140k. One data point, not a benchmark. But note that it agrees with the SWE-Bench Pro chart, not the Terminal-Bench one.

One bright spot: long-form writing. Multiple testers independently report that 5.6's prose structure and stylistic control are markedly improved, needing minor edits instead of reconstruction. In a release cycle where every lab is betting the farm on coding, shipping a better writer is unexpected.

The Theo protocol

Theo (t3.gg) has given one of the most useful accounts of this launch cycle. When GPT-5.5 dropped, he didn't like it, said so publicly, then spent two months re-prompting and reconfiguring before concluding he couldn't use anything else for code. His conclusion was that prompts are technical debt: every carefully tuned setup silently degrades when the model underneath it changes, and an apparently worse model often reflects an old, unmigrated configuration rather than a real capability drop.

Apply that here and the day-one GPT-5.6 takes, including the glowing ones, are mostly measurements of people's old prompts. The honest evaluation window hasn't closed yet.

Theo is also already doing the unglamorous work on 5.6: he surfaced the bug where setting Sol to Ultra forces every sub-agent to inherit Ultra with no per-agent override, burning tokens for nothing. OpenAI's response amounted to "daily use of Medium is sufficient," which reads as an acknowledgment that the Ultra tier's cost is hard to justify. He distrusts Frontier Codebench because Opus 4.8's scores across reasoning levels look like a random number generator. He's building his own benchmark framework because the existing ones annoy him. That is the posture this blog aims for, and his sources remain the closest available substitute for running these evals independently.

How to actually use each model

• Orchestration, long chains, deep review, anything a human will read: Fable 5. The SWE-Bench Pro gap is real, the state-tracking edge over long agent chains is real, and the writing-quality gap on human-facing output is consistent. This is what the extra $5/$25 per million buys.
• Checklist-heavy autonomous agent work, including browsing, OS tasks, and terminal grinding: GPT-5.6 Sol at Medium effort. Its Terminal-Bench and BrowseComp numbers reflect something true: it stays oriented and doesn't quit. Never use Ultra (inheritance bug), and never use Full Access on a machine you care about.
• Default 5.6 tier for everyday work: Terra. It sits within 2 to 3 points of Sol on most suites at half the price, though field reports say the real-world savings are smaller than the pricing page implies. Test on your own workload before believing either number.
• Long-form drafting: GPT-5.6, unironically. The writing improvement seems like one of the more under-reported parts of this launch.
• Cheap, fast, good-enough: Grok 4.5 offers sufficient capability at a price and speed that make retry loops rational, which makes it an easy recommendation this month. Luna if you're staying in the OpenAI stack.
• Judging any of this from launch-day takes: re-prompt for the new model, give it two weeks, then decide, per the Theo protocol above.