OpenAI released GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output tokens, half the list price of GPT-6.1 Sol. Nineteen days earlier, Anthropic shipped Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens, a tenth of Haiku 4.5's rates. These are the models meant to run subagents, routing and computer use, the busiest parts of an agent system. October's contest was over what a dollar buys, not who tops the chart.
Two rate cards, nineteen days apart
Up to 100,000 tokens, Haiku 5.5 costs $0.10 input, $0.50 output and $0.01 for cache reads; above that, every rate rises fivefold, to $0.50, $2.50 and $0.05. Haiku 4.5 was $1 and $5; Sonnet 5.5 is $2 and $10. Anthropic puts the average saving at about 75%, after a new tokenizer that spends slightly more tokens per task, and says around 90% of Haiku 4.5 requests fit under the threshold.
Terra lists at $1 input, $0.05 cached and $5 output, with a 1,050,000-token window and up to 128,000 tokens of output; prompts over 272,000 tokens pay double for input and caching and 1.5 times for output, across the whole request. GPT-6.1 Sol costs $2, $0.10 and $10. Haiku 5.5 takes up to 1 million tokens in and 128,000 out.
Neither is the cheapest model on the market. GPT-6 Luna has listed at $0.10 and $0.50 since September 22, Haiku 5.5's exact base rate, and Terra costs what Haiku 4.5 did. What fell twice in October is the price of a given level of capability.
Haiku 5.5 runs on Anthropic's API, AWS, Google Cloud and Microsoft Azure, and in Claude.ai's free and paid plans and Claude Code. Terra went generally available in GitHub Copilot the same day on paid plans, billed at OpenAI's list price.
What each lab says it bought
By Anthropic's own table, Haiku 5.5 scores 72.4% on the offline subset of OSWorld 2.1 and 39.2% on Terminal-Bench 4.0, against 15.7% and 0.0% for Haiku 4.5 and 48.9% and 16.4% for GPT-6 Luna. Sonnet 5.5 leads at 83.9% and 70.6%, and Anthropic concedes that Sonnet and Opus "remain better choices for complex agentic coding." Haiku's pitch is the subagent that works under them.
OpenAI's claims are harder to check. Its announcement declined our fetch, so its ChatGPT and Codex details and its own benchmark figures, seen only through an aggregator, are left out. On the record: the system card addendum says Terra shares GPT-6 Astra's data and training and, like Astra, is treated as Critical in cybersecurity, and GitHub's early testing found it solved "noticeably more terminal and C# coding tasks" than GPT-5.6 Terra. Haiku 5.5's safeguards still block penetration testing. Small models now ship with big-model guardrails.
A tenfold price gap, a one-cent cost gap
Artificial Analysis, which runs both on its own harness, complicates the sticker. At Terra's high effort and Haiku's xhigh, a task on its index costs $0.26 and $0.25. Haiku charges a tenth as much per token but emitted 136,000 output tokens per task to Terra's 17,000, and scored 34 to Terra's 42. Terra led on Terminal-Bench 4.0 (42% to 29%) and AutomationBench (58% to 36%); Haiku led AA-Briefcase, a knowledge-work test, at 1,524 Elo to 1,264, and streamed 191 tokens per second to Terra's 91.
At maximum effort, Terra scores 51 at $0.96 per task; Haiku scores 36 at $0.41, the most verbose of 22 comparable models. On the Vals Index, Terra ranks 11th at 60.47% and Haiku 14th at 58.73%, and Haiku costs more per test, $3.17 to $2.92.
Each has a rival on its flank. Against Luna at the same price, Haiku 5.5 leads Artificial Analysis's index 36 to 28 and Terminal-Bench 32.8% to 12.6%, though Luna leads AutomationBench, 53.2% to 35.4%, a score Artificial Analysis said at launch was likely understated by a safety-refusal issue. Against Sol at twice the price, Terra at max effort matches Sol's 56% on Terminal-Bench 4.0 (Terra, Sol), scores 51 to Sol's 55, and saves about a quarter per task, $0.96 to $1.27, rather than half.
How the agent budget changes
Three habits follow. Budget per task, not per token: effort is now a pricing dial, and Haiku 5.5 alone spans $0.04 a task at low effort to $0.41 at max on Artificial Analysis's runs. OpenAI's guide says to start at medium and raise effort only when the quality is worth the tokens, now a finance memo as much as an engineering note.
Watch the thresholds. Haiku 5.5's 100,000-token line is applied per request and counts cache reads and writes, so an agent loop whose context creeps past it pays five times as much on every rate, cached tokens included. Terra's surcharge starts at 272,000. Compaction used to be a quality measure; it is now a cost control, and the sharpest one on the cheaper model.
Route by latency as well as price. At the settings above, Terra's first token arrived in 8.2 seconds to Haiku's 45.7, though Haiku streams twice as fast once started. A subagent that keeps a user waiting and one that runs overnight want different models, possibly from different vendors.
The 30,000-foot read
The list prices favor Haiku; the capability scores favor Terra. Per task, the two meet at nearly the same cost by opposite roads, a cheap model that thinks at length and a dearer one that thinks briefly; at that meeting point Terra scored higher and Haiku wrote faster. Neither lab decides which of those your workload rewards. You do, through effort settings, context discipline and routing, and the invoice will track those choices more closely than any rate card.
Frontier peaks still sell launch posts, but in a well-built agent system the volume runs through models like these, and both labs now price them as the main event. The next round will be fought over cost per completed task, ideally measured by someone other than the vendor. Until then, read the token counts before the price sheet.
Sources: OpenAI API changelog (Terra's October 26 release; Luna's September 22 release); GPT-6.1 Terra model page (prices, long-context surcharge, window, output limit; source of an image reference); OpenAI API pricing (Terra, Sol and Luna prices; chart); Using GPT-6 (effort guidance); GPT-6.1 Terra system card addendum (shared training with GPT-6 Astra, Critical cybersecurity treatment); GitHub changelog (Copilot general availability on Pro, Pro+, Max, Business and Enterprise plans, list-price billing, early-testing quote); Introducing Claude Haiku 5.5 (October 7 release, prices, 75% average saving and tokenizer, 90% of requests, benchmark table, "remain better choices" quote, subagent positioning, safeguards, platforms; source of an image reference); Claude API pricing (Haiku 5.5 tiers and threshold rules, Haiku 4.5 and Sonnet 5.5 prices; chart); The Batch, issue 375 (Haiku 5.5 input and output limits, Claude.ai and Claude Code availability); Artificial Analysis: Haiku 5.5 launch (AutomationBench caveat); Artificial Analysis: Terra (high) vs Haiku 5.5 (xhigh) (cost per task, tokens, scores, speed, time to first token); Artificial Analysis: GPT-6.1 Terra and Claude Haiku 5.5 (max-effort scores, cost per task, verbosity); Artificial Analysis: Haiku 5.5 vs GPT-6 Luna (scores and cost per task by effort); Artificial Analysis: Terra vs Opus 5.5 (Terra's Terminal-Bench 4.0 at max); Artificial Analysis: Argon vs GPT-6.1 Sol (Sol's Terminal-Bench 4.0); Artificial Analysis: Haiku 5.5 vs GPT-5.6 Terra (GPT-5.6 Terra's Terminal-Bench 4.0, footer only); Artificial Analysis Intelligence Index v5 (Sol's score and cost per task); Vals Index (v2.2 ranks and cost per test, October 27). CAVEATS: The OSWorld 2.1 and Terminal-Bench 4.0 figures for Haiku 5.5, Haiku 4.5, Luna and Sonnet 5.5 are Anthropic-reported and run higher than Artificial Analysis's own measurements (Haiku 5.5 39.2% vs 32.8%; Luna 16.4% vs 12.6%); Artificial Analysis and Vals figures are independent. All Artificial Analysis scores here come from its Intelligence Index v5.0, published October 27, after Terra's launch, and are not comparable with launch-day v4.3 scores; its speeds are snapshots that vary by day. OpenAI's launch post declined our fetch, so ChatGPT and Codex availability and OpenAI's own benchmark figures are not verified here; we did not route around the refusal. The chart shows base output rates for short prompts: Haiku 5.5 charges $2.50 for prompts over 100,000 tokens, and OpenAI's output rates rise 1.5 times above 272,000. "About a quarter" is our arithmetic from Artificial Analysis's per-task costs. CORRECTIONS to our leads: Anthropic's OSWorld 2.1 figure is for the offline subset; Haiku 5.5 matched, rather than undercut, GPT-6 Luna's $0.10/$0.50 list price, so the price floor fell in capability-adjusted terms, not absolute ones; the lead's Artificial Analysis scores (48 for Terra, 43 for Haiku) are launch-era v4.3 figures, now 51 and 36 on v5.0, and the 48 could not be confirmed at Artificial Analysis; the 237 tokens per second is The Batch's launch-time report of Artificial Analysis's measurement (its page now shows 241 at max effort and 191 at xhigh); the Terminal-Bench 4.0 claim of 50.3% for Terra vs 21.5% for GPT-5.6 Terra appears only in aggregator coverage and is not used (Artificial Analysis measures 56% and 35% at max effort); Terra's 1M context per Artificial Analysis is 1,050,000 tokens in OpenAI's docs.