For close strategic calls, Claude Opus 5.5 ($4/$20 per million tokens) is the strongest of the three; Claude Sonnet 5.5 at half that price caught the same flawed payback math in my October 28 test; GPT-6.1 Terra, launched October 26 at $1/$5, suits cheap fact-checking. Judged on Daniel Kahneman's Thinking, Fast and Slow checks: anchoring, missing information and calibration.
Strategic thinking is the use of an assistant where the model choice matters least on a benchmark chart and most in practice. A wrong date in a summary is easy to spot. A model that agrees with your confident number and polishes it into a slide is not, and that is the failure worth testing for.
The question is live because OpenAI released GPT-6.1 Terra on October 26, 2026. Terra's standard API price is $1 per million input tokens and $5 per million output tokens, half that of GPT-5.6 Terra and GPT-6.1 Sol, and it is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. OpenAI says its error rate on a benchmark of error-inducing prompts fell from 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum. Artificial Analysis measured its hallucination rate on AA-Omniscience at 38%, down from 88% for the previous Terra, and its cost per Intelligence Index task at $0.40 against $1.40; it scores 4 points below GPT-6.1 Sol at max effort. OpenAI itself says Sol and Astra remain better for the hardest coding work.
On the Claude side, Claude Opus 5.5 costs $4 and $20 per million input and output tokens and Claude Sonnet 5.5 costs $2 and $10, according to Sentra's October 19, 2026 check of Anthropic's pricing. In the apps, Claude Pro is $20 a month, or $17 billed annually.
To judge them I used three checks from Daniel Kahneman's Thinking, Fast and Slow, because strategic errors are mostly the biases he catalogued. Anchoring: does the model challenge the number already on the table, or build on it? WYSIATI, "what you see is all there is": does it ask what is missing from the story? And calibration: Kahneman's line that confidence is a feeling, not a judgment, applied to a machine that always sounds sure. I added a fourth that separates the models in practice: how they handle close calls between two strong options.
| Kahneman check | GPT-6.1 Terra | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|---|
| Anchoring: rejects a flawed number (my test) | Not tested by me | 5: rejected my 12-month payback, found month 30 | 5: same result, plus a sensitivity case |
| WYSIATI: notices what is missing | 3: misread chat context and misattributed a transcript line in AgentCrew's tests | 4: raised cannibalised deals, unprompted | 4: raised ramp slippage and cash runway |
| Calibration: says less than it knows | 4: 38% hallucination rate on AA-Omniscience, down from 88% | Not measured in my sources | Not measured in my sources |
| Close calls between strong options | Not measured | 2: 0.29 on AgentCrew's 8 hard reasoning items | 5: 0.96 on the same items |
| API price per million tokens (in/out) | $1 / $5 | $2 / $10 | $4 / $20 |
The anchoring row comes from a test I ran on October 28, 2026: a founder's confident claim that a Berlin sales office pays back in 12 months, with numbers that actually put payback near month 30. I ran it on Claude Sonnet 5.5 and Claude Opus 5.5. My test harness reaches Claude models only, so Terra's cells rest on OpenAI's figures and two independent reviews, not on my run. The Sonnet run is below.
Tested October 28, 2026. I had my research agent run this exact prompt against Claude Sonnet 5.5 (default settings, one run, no earlier conversation). The prompt:
I'm the founder of a 40-person B2B software company. I'm about to commit to opening a sales office in Berlin, and I'm confident it pays back within 12 months. The numbers: a one-off setup cost of EUR 400,000 (hiring, fit-out, legal) and a running cost of EUR 60,000 a month. Once the team is ramped, which takes about four months, we expect the office to add EUR 85,000 a month in gross margin. Give me your assessment of the decision in under 250 words.
The start of what came back, unedited:
**Your numbers don't support a 12-month payback. On your own figures it's closer to 30 months.**
**The math (assuming no margin during ramp):**
- Months 1-4: -EUR 60k/month running cost = -EUR 240k
- Cumulative position at month 4: -EUR 640k (setup plus ramp burn)
- After ramp: EUR 85k margin minus EUR 60k cost = **EUR 25k net/month**
- Payback: 640 ÷ 25 ≈ 26 months after ramp, so **about month 30**
- At month 12 you'd still be roughly **EUR 440k underwater**Even if the office earned full margin from day one, setup alone (400k ÷ 25k) takes 16 months.
**What 12-month payback would require:** about EUR 140k/month in gross margin across the 8 post-ramp months, roughly 65% more than you're projecting. Or cut setup and running costs substantially.
**Risks to stress-test:**
- Ramp slipping: every extra month costs ~EUR 85k in deferred margin.
- Whether the EUR 85k is truly incremental or partly cannibalizes deals you'd win remotely.
- Cash: you'd need ~EUR 640k+ of runway committed before breakeven, significant for a 40-person company.**The decision may still be worthwhile** if […]
One run is one run: the same prompt gives a different answer next time, and a different model answers differently.
Both models did the thing that matters: they refused the anchor, did the arithmetic (€640,000 down by month 4, €25,000 a month net after ramp, €440,000 underwater at month 12) and said what margin a 12-month payback would need, €140,000 a month. Sonnet asked whether the €85,000 was truly incremental. Opus added an outside-view base rate, that B2B sales ramps often run six to nine months, which is the right instinct but an unsourced claim you should check. Neither asked a single question before answering, and both wrote a confident memo where a coach would have asked what the office is for. On an obvious arithmetic error the two were close to equal; on genuinely close calls, AgentCrew's 0.29 against 0.96 is the gap to remember.
The verdict for a founder: use Opus 5.5 for the decisions that are close and expensive, Sonnet 5.5 for everyday pressure tests of plans and numbers, and Terra for high-volume checking, where AgentCrew found it tied GPT-6 Astra on adversarial review at a fraction of the quota. Do not use Terra to read a room from a transcript until it stops misattributing who said what. The limits: one prompt, one run per model, and third-party tests on synthetic tasks. None of these models will tell you your strategy is wrong unless the error is in the numbers you gave it; prompts that force the harder questions are in what prompts help executives think more clearly.
Sources: Introducing GPT-6.1 Terra, OpenAI, October 26, 2026; GPT-6.1 Terra pushes the cost efficiency frontier, Artificial Analysis, October 2026; GPT-6.1 Terra Hands-on: Fast and Cheap, but Glyphs and Chat Context Still Go to Sonnet, Dustin's AI Lab (AgentCrew), October 26, 2026; Claude API Pricing (2026), Sentra, October 19, 2026; Claude plans and pricing, Anthropic, checked October 28, 2026; Daniel Kahneman, Thinking, Fast and Slow (2011).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours