FULL STORY

GPT-6.1 Terra: Budget Launch Meets Real-World Benchmarks

OpenAI launched GPT-6.1 Terra at half the price of GPT-5.6. Early benchmarks and user tests showed mixed results, with spatial reasoning still trailing Astra despite its low cost.

2026-10-27 ~ 2026-10-29 · 4 episodes · 27 posts

Episode 1 · GPT-6.1 Terra nears Sol at half cost but Astra leads spatial reasoning (2026-10-27, 3 posts)

GPT-6.1 Terra scored 8% on MazeBench, nearly matching Sol's 9% at half the cost, but GPT-6 Astra still holds SOTA in spatial reasoning tasks like Gems-Moves.

Episode 2 · GPT-6.1 Terra Tested: Fast and Cheap, But Not Smarter (2026-10-28, 11 posts)

OpenAI released GPT-6.1 Terra (max tier, focused on computer use) priced at $1/M input and $5/M output—half of GPT-5.6—with an official claim of being 9x cheaper per task. Community benchmarks quickly piled up; the consensus is clear speed and cost advantages, but a capability ceiling below the Opus tier—"cheaper, not better" is the dominant verdict.

Confirmed

  • Pricing: $1/M input, $5/M output, half of GPT-5.6; @aigenerator2077 broke down the "9x cheaper" claim as a mix of token-efficiency factors, but the math doesn't fully add up (relayed in m11).
  • Independent tests show cost/speed wins: on the same rainbow jelly-bear task Terra cost $4.15 in 25 minutes vs Opus 5.5's $13.16 in 40 minutes (m3); @VraserX reports a coding task done 3x faster at roughly one-ninth the cost, nearly matching Opus (m5); @robocurve reports Terra beating Opus 5.5 at controlling a real robot, a 3.7x improvement over three months (relayed m1).
  • Artificial Analysis's intelligence-vs-cost-per-task scatter (78 models) places GPT-6.1 Terra medium/high on the Pareto frontier in the most attractive quadrant; OpenAI models occupy nearly the entire frontier, from $0.006/task entry models up to $1+ (m4, m6).
  • Speed: Terra leads Sol decisively (91 vs 61-81 tokens/s, via m7); @chongdashu's coding test finds capability clearly below Sol/Astra/Opus tier but extreme speed—tasks finished before he could switch sessions (m9).
  • Negatives: Reddit user bofors14 calls terra's coding poor across the board—as a sub-agent it only burns more tokens than luna max with no gain, and it disappoints as a main agent (m8); Towards AI's ToneBench writing eval (10 YouTube-script tasks × 5 drafts, 4 LLM blind judges) ranked Terra #23, concluding "cheaper, not better" (m10).

Unconfirmed

  • All cost and performance figures come from personal or third-party self-built benchmarks, not official ones, and are highly scenario-dependent; @IamNotGorbachev notes split community verdicts ("useless," "terrible," "powerful" coexist)—your mileage varies (m2).
  • @robocurve's robot-control data is a secondhand reference; original details unverifiable.

Why it matters

Terra embodies OpenAI's strategy of capturing the enterprise agent/computer-use market with a cheap, fast model: on the intelligence-cost curve, OpenAI nearly monopolizes the Pareto frontier, undercutting rivals' value narratives. For developers, though, it means weighing "good enough and cheap" against top-tier capability—it suits high-frequency, cost-sensitive tasks, not the complex work reserved for the Opus tier.

Episode 3 · Cognition's FrontierCode 2.0 leaderboard lands; hidden model Obsidian 3 tops Opus 5.5 (2026-10-28, 11 posts)

Cognition's FrontierCode 2.0 benchmark is now live on Artificial Analysis: Claude Opus 5.5 leads at 46.4%, the top 20 is dominated by OpenAI and Anthropic models, and Chinese models sit at the bottom. A hidden model, Obsidian 3, appears to have leaked in the eval data with 46.8%, surpassing the leader and widely suspected to be the unreleased Fable 5.5.

Confirmed

  • Per @bookwormengr, the benchmark comprises 132 tasks on real open-source repos, scored by whether maintainers would merge the changes, with tasks kept private to prevent contamination.
  • Head results: Claude Opus 5.5 tops at 46.4% (Opus 5.5 max shown at 45.5% in charts cited by @brandongalang and @YashasGunderia); Kimi and DeepSeek are close behind (@bookwormengr).
  • Per @scaling01, the top 20 is almost entirely OpenAI and Anthropic; only Grok 4.7 xhigh at 28.6% (No. 20) breaks through, with Chinese models at the bottom.
  • OpenAI's GPT-6.1 Terra ranks highly at low pricing: $1/M input, $5/M output, $0.05 cached input (@bookwormengr).
  • Opus 5.5 [high] scores 34.5% at $2.42 per rollout (@teortaxesTex).
  • Citing 1337hero, @bookwormengr notes Anthropic models burn 200%-300% more tokens, take 2x longer, and cost 10-50x more on such frontier tasks.

Unconfirmed

  • Hidden model Obsidian 3: reportedly 46.8% at high effort, beating Opus 5.5 max's 45.5% (@kimmonismus, @danielmac8, @YashasGunderia); @teortaxesTex estimates its cost at 3.6x Opus 5.5, while @danielmac8 claims it is actually cheaper. @YashasGunderia notes it runs on Claude-specific infrastructure. Identity, cost and data authenticity are unconfirmed.

Why it matters

  • Merge-based scoring on hidden tasks makes this a hard-to-game, credible measure of agentic coding, offering a clear snapshot of the frontier landscape.
  • @brandongalang reads it as the clearest rebuttal of claims that Chinese open models threaten Anthropic/OpenAI — though 1337hero's numbers show intelligence leadership doesn't imply efficiency leadership. A real Obsidian 3 would mean an unreleased model already matches or beats the best, pending official confirmation.

Episode 4 · GPT-6.1 Terra Ranks 8th on Agent Arena at $0.62 Median Cost per Task (2026-10-29, 2 posts)

GPT-6.1 Terra (Max) ranks 8th on LMArena's Agent Arena based on nearly 9,000 real agentic sessions, with a median cost of just $0.62 per task and a net improvement of +10.