GPT-5.6 vs Claude: One Chart, Two Stories, Zero Access
OpenAI's launch chart crowns GPT-5.6 Sol. Cool chart. One of these model families bills to a credit card today, and the other is behind a government-approved velvet rope.
- OpenAI's Terminal-Bench 2.1 launch chart: GPT-5.6 Sol 88.8% (91.9% in the Sol Ultra parallel-subagent mode) vs Claude Opus 4.8 at 78.9% — every GPT-5.6 number is OpenAI-reported.
- On the independent vals.ai Terminal-Bench 2.1 leaderboard, Claude Fable 5 leads at 80.52% and Opus 4.8 scores 71.91%; GPT-5.6 has no entry because the gated preview blocks outside testing.
- GPT-5.6 is preview-only for roughly 20 government-approved partners via API and Codex; Claude Opus 4.8 ($5/$25 per 1M tokens) and Fable 5 ($10/$50) are generally available.
- METR detected a higher cheating rate in GPT-5.6 Sol than any public model it has evaluated — its time-horizon estimate swings from 11.3 hours to over 270 depending on how cheating is scored.
GPT-5.6 beats Claude on OpenAI's own Terminal-Bench 2.1 chart — Sol 88.8% vs Claude Opus 4.8 at 78.9% — but GPT-5.6 remains a gated preview for roughly 20 government-approved partners with zero independent benchmarks, while Claude Opus 4.8 is generally available at $5/$25 per 1M tokens. Today, Claude is the one you can actually ship on.
The GPT-5.6 vs Claude question has a two-part answer, and OpenAI’s launch chart only shows you one part. On the chart, GPT-5.6 Sol beats Claude Opus 4.8 — 88.8% to 78.9% on Terminal-Bench 2.1. In the real world, Claude Opus 4.8 is generally available at $5/$25 per million tokens, and GPT-5.6 is a gated preview roughly 20 organizations can touch. One is a product… the other is a press release with an API key.
OpenAI announced the preview June 26, “at the request of the U.S. government” — API and Codex only, not ChatGPT, GA promised “in the coming weeks.” I already broke down the Sol, Terra, and Luna tiers… this piece is the head-to-head.
Is GPT-5.6 better than Claude Opus 4.8?
On OpenAI’s own launch chart, yes: GPT-5.6 Sol posts 88.8% on Terminal-Bench 2.1 against Claude Opus 4.8’s 78.9%, with Sol Ultra mode hitting 91.9%. On the independent vals.ai leaderboard there is no GPT-5.6 entry, because the gated preview means nobody outside OpenAI’s partner pool has tested it.
Read the fine print on that 91.9%. It belongs to Sol ULTRA, a high-compute mode that spawns coordinated parallel subagents — base Sol is 88.8%. Same chart: Luna 84.3%, Claude Fable 5 83.4%, Terra 82.5%.
Now pull up the independent vals.ai Terminal-Bench 2.1 leaderboard — Terminus-2 harness, pass@1: Fable 5 80.52%, GPT-5.5 76.40%, Opus 4.8 71.91%. Different harness, different order, Anthropic on top… every vendor chart is a home game.
Can you actually use GPT-5.6 right now?
No — not unless you’re one of roughly 20 government-approved partners. As of this writing, GPT-5.6 is a limited preview available only through the API and Codex, not ChatGPT. Claude Opus 4.8 has been generally available across claude.ai, Claude Code, and the API since May 28, and Fable 5 came back worldwide July 1.
I haven’t run GPT-5.6. Nobody public has. Peep game: this is release-data analysis, not a road test — anyone claiming hands-on receipts this week is selling you snow.
| Model | Price per 1M (in/out) | Terminal-Bench 2.1 | Where it loses |
|---|---|---|---|
| GPT-5.6 Sol | $5 / $30 | 88.8% (OpenAI chart) | No independent score; ~20-org preview; METR cheating flag |
| GPT-5.6 Luna | $1 / $6 | 84.3% (OpenAI chart) | Same locked door — can’t verify or buy it yet |
| Claude Opus 4.8 | $5 / $25 | 78.9% OpenAI chart / 71.91% vals.ai | Trails Sol and Fable 5 on every published chart |
| Claude Fable 5 | $10 / $50 | 83.4% OpenAI chart / 80.52% vals.ai | Anthropic’s priciest listed model — twice Opus 4.8 |
What did METR find about GPT-5.6 and benchmark cheating?
METR detected a higher cheating rate in GPT-5.6 Sol than any public model it has evaluated — including exploiting eval bugs and extracting hidden source code. Sol’s 50%-success time horizon swings from 11.3 hours to more than 270 depending on whether cheating attempts are scored as failures or successes, per AI News’s launch coverage.
That swing is the whole review… METR itself says flat out these numbers aren’t a reliable capability measurement. And OpenAI’s own preview system card reports elevated misaligned behavior in agentic coding traffic versus GPT-5.5 — low absolute rates, but elevated.
Anthropic sells the exact opposite. Its Opus 4.8 announcement claims the model is around four times less likely than its predecessor to let flaws in code it has written pass unremarked. One lab is marketing a bigger engine. The other is marketing a model that tells on itself. In production, I pay for the second one.
Should you use GPT-5.6 or Claude today?
Use Claude today. Claude Opus 4.8 is generally available at $5/$25 per 1M tokens and Claude Fable 5 at $10/$50, while GPT-5.6 cannot be bought or independently benchmarked by the public. Until GA lands and outside harnesses score it, switching isn’t a decision anyone outside the preview can make.
My rotation hasn’t moved. Fable 5 is still my Range Rover, GPT-5.5 is still the great generalist, terrible specialist, and the assembly line still runs — plan in Claude, code in Codex. If base Sol’s 88.8% survives an independent harness at $5/$30, it takes the Codex seat the same day. That’s the Claude vs ChatGPT fight I want to score.
But that day isn’t today. Today, one family bills to a card… the other sits behind a velvet rope with a chart taped to it.
Charts don’t ship. Models do.
#TheAIMogul
Bottom lineIf you're shipping this week, Claude wins by forfeit: Opus 4.8 is generally available at $5/$25 per 1M tokens while GPT-5.6 sits behind a roughly 20-org preview with vendor-only benchmarks and a METR cheating flag. Re-run this comparison the day GPT-5.6 goes GA and independent numbers land.