OpenAI GPT-6.1 Terra vs Claude Sonnet 5.5, Gemini 4 Argon and Grok 4.7: API pricing, context and computer use compared

OpenAI released GPT-6.1 Terra on October 26, 2026 as the efficient tier of its GPT-6 family, priced at $1.00 per million input tokens and $5.00 per million output tokens in the API. The launch itself, along with Terra's availability in ChatGPT Work, Codex and GitHub Copilot and its safety results, is covered in our launch report on GPT-6.1 Terra. This article looks at a different question that developers and procurement teams face the same day: where Terra sits against the models it will be tested against, namely Anthropic's Claude Sonnet 5.5 and Claude Haiku 5.5, Google's Gemini 4 Argon and Gemini 3.8 Flash, and xAI's Grok 4.7.

The comparison relies only on figures each provider publishes in its official API documentation and pricing pages. Prices are as listed on the official pricing pages and model pages of OpenAI, Anthropic, Google and xAI on October 26, 2026, in US dollars per million tokens for standard, non-regional processing. The short version is that Terra lands at exactly half the list price of Claude Sonnet 5.5 and of Gemini 4 Argon during its introductory period, below Grok 4.7 on input but close to it on output, and well above the two budget models, Claude Haiku 5.5 and Gemini 3.8 Flash. Its strongest structural advantage is a very cheap cached-input rate combined with a hosted computer-use tool, while its weakest point on paper is a long-context surcharge that some competitors do not apply.

These are not like-for-like products in every respect. Gemini 4 Argon is Google's flagship rather than a mid-tier model, Grok 4.7 is the only general language model xAI lists on its models page, and Claude Haiku 5.5 is positioned by Anthropic for high-volume classification, routing and subagent work. The pairing reflects price brackets and typical use, not equivalent capability, and no provider has published scores that allow a direct, same-conditions benchmark comparison across all six models.

··········

LIST PRICES FOR GPT-6.1 TERRA AND ITS COMPETITORS.

The table below shows the standard list prices, the cached-input or cache-read rate, the batch discount and the rule each provider applies to long prompts. Cache-write charges, which differ in structure between providers, are discussed separately below.

........

Model	Input	Cached input	Output	Batch	Long-prompt rule
GPT-6.1 Terra	$1.00	$0.05	$5.00	50% off	Over 272K input: 2x input and cache, 1.5x output
Claude Sonnet 5.5	$2.00	$0.10	$10.00	50% off	Flat price up to 1M tokens
Claude Haiku 5.5	$0.10	$0.01	$0.50	50% off	Over 100K prompt: $0.50 / $0.05 / $2.50
Gemini 4 Argon	$2.00	$0.10	$10.00	$1.00 / $5.00	No long-context tier listed
Gemini 3.8 Flash	$0.75	$0.075	$3.75	$0.375 / $1.875	No long-context tier listed
Grok 4.7	$2.00	$0.50	$6.00	20% off	200K+ prompt: $4.00 / $1.00 / $12.00

........

Two of the Google prices are temporary, and that matters for anyone budgeting beyond the next few weeks. Gemini 4 Argon's $2.00 and $10.00 are introductory rates that Google lists as valid through November 30, 2026; from December 1, 2026 the same model is listed at $4.00 for input, $20.00 for output and $0.20 for cached input. Gemini 3.8 Flash is listed at $0.75 and $3.75 through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027. Once both increases take effect, Terra will cost a quarter of Argon per token and will be cheaper than Gemini 3.8 Flash on both input and output.

The cached-input column explains much of the cost picture for agents. OpenAI bills Terra's cached input at 5% of the uncached rate, the same ratio Anthropic now applies to Sonnet 5.5 and Google applies to Argon. Grok 4.7's cached input is $0.50, a quarter of its $2.00 input price, which makes repeated context comparatively expensive on xAI's model. Claude Haiku 5.5 is the cheapest model in the group on every line, and Anthropic's cache-read rate for it is $0.01 per million tokens for prompts up to 100,000 tokens.

Cache writes follow different models. OpenAI charges $1.25 per million tokens to write Terra's prompt cache with 30-minute retention and $2.50 with 12-hour retention. Anthropic charges $2.50 and $4.00 for Sonnet 5.5's 5-minute and 1-hour cache writes, and $0.125 and $0.20 for Haiku 5.5 on short prompts. Google lists Argon's caching as implicit only, with no separate storage price on the pricing page, while Gemini 3.8 Flash also carries a storage fee of $0.50 per million tokens per hour through December 31. xAI does not list a cache-write charge for Grok 4.7.

··········

CONTEXT WINDOWS, OUTPUT LIMITS AND COMPUTER USE.

Terra's other headline feature is computer use, so the comparison has to include which competing models can operate a screen through the provider's own API tooling. The table summarizes what each official model page and tool page states.

........

Model	Context window	Max output	Computer use via API	Reasoning control
GPT-6.1 Terra	1,050,000	128,000	Yes, hosted tool	6 effort levels, none to max
Claude Sonnet 5.5	1M	128K	Yes, computer toolset	Adaptive, default high
Claude Haiku 5.5	1M	128K	Yes, computer toolset	Adaptive, default medium
Gemini 4 Argon	1,048,576 shared	65,536 default	Yes	low, medium, high
Gemini 3.8 Flash	1,048,576 input	65,536	Yes, preview	low, medium, high
Grok 4.7	500,000	Not stated	Not listed	low to xhigh, default high

........

All six models handle long inputs, but the details differ. Terra, Sonnet 5.5 and Haiku 5.5 each accept around one million tokens with up to 128,000 output tokens, and Anthropic offers up to 300,000 output tokens on its Batch API with a beta header for both Claude models. Google's Gemini 4 Argon uses a unified 1,048,576-token window shared by input, thinking and output, with a default output cap of 65,536 tokens that can be raised with background execution. Grok 4.7 has the smallest window in the group at 500,000 tokens.

On computer use, OpenAI exposes the capability as one of the hosted tools in its Responses API for Terra. Anthropic's current `computer_toolset_20260801` supports Sonnet 5.5 and Haiku 5.5 without a beta header, although Anthropic notes that declaring the toolset adds about 4,500 input tokens per request and that screenshots are billed as image input. Google's computer use documentation lists both Gemini 4 Argon and Gemini 3.8 Flash and describes browser, Android-optimized mobile and desktop environments, while the Gemini 3.8 Flash model page marks computer use as a preview capability. xAI's pricing page lists web search, X search, code execution, image generation, file attachments, collections search, image and video understanding and remote MCP as server-side tools for Grok, and does not list a computer use tool.

Knowledge cutoffs are only partly comparable. OpenAI gives June 1, 2026 for Terra, and Anthropic gives June 2026 for both Sonnet 5.5 and Haiku 5.5. The Google and xAI pages consulted for this comparison do not state a cutoff for Argon, Gemini 3.8 Flash or Grok 4.7.

··········

WHAT THREE TYPICAL WORKLOADS COST ON EACH MODEL.

List prices only become useful when applied to a workload. The three scenarios below use the standard rates from the tables above, ignore cache-write charges and taxes, and assume that each model consumes the same number of tokens for the same task. That last assumption is the weakest one: each provider uses its own tokenizer, and Anthropic itself notes that Haiku 5.5's newer tokenizer produces roughly 30% more tokens than Haiku 4.5's for the same text. Real costs will also depend on how many reasoning tokens each model generates, which is billed as output on every platform listed here.

The first scenario is a customer-support assistant handling 20,000 requests per day for 30 days. Each request sends 6,000 input tokens, of which 5,000 are a cached system prompt and knowledge block and 1,000 are new, and returns 400 output tokens. The second is an agentic computer-use or coding session of 40 turns, where each turn sends 80,000 input tokens, 70,000 of them cached, and produces 1,500 output tokens. The third is a long-document job: 100 documents of 400,000 tokens each, with no caching and 4,000 output tokens per document.

........

Model	Support bot, per month	Agent session, 40 turns	100 long documents
GPT-6.1 Terra	$1,950.00	$0.84	$83.00
Claude Sonnet 5.5	$3,900.00	$1.68	$84.00
Claude Haiku 5.5	$210.00	$0.10	$21.00
Gemini 4 Argon, to Nov 30	$3,900.00	$1.68	$84.00
Gemini 4 Argon, from Dec 1	$7,800.00	$3.36	$168.00
Gemini 3.8 Flash, to Dec 31	$1,575.00	$0.74	$31.50
Gemini 3.8 Flash, from Jan 1	$3,150.00	$1.47	$63.00
Grok 4.7	$4,140.00	$2.56	$164.80

........

For the support assistant, each Terra request costs 1,000 × $1.00 plus 5,000 × $0.05 plus 400 × $5.00, divided by one million, or $0.00325, which over 600,000 requests comes to $1,950. The same arithmetic gives $0.0065 per request on Sonnet 5.5 and on Argon at its introductory price, and $0.0069 on Grok 4.7, whose higher cached-input rate cancels most of the advantage of its lower output price. Haiku 5.5, at $0.00035 per request, is cheaper than Terra by a factor of about nine in this scenario.

The agent session follows the same pattern. A Terra turn costs 10,000 × $1.00 plus 70,000 × $0.05 plus 1,500 × $5.00, or $0.021, and 40 turns cost $0.84. Sonnet 5.5 and introductory Argon double that to $1.68, and Grok 4.7 reaches $2.56 because 70,000 cached tokens per turn are billed at $0.50 rather than $0.05 or $0.10. On Anthropic's models the roughly 4,500-token overhead of the computer toolset would add to every turn, and on all platforms screenshots add image input that this simplified model does not price separately.

The long-document job is where Terra's surcharge appears. A 400,000-token prompt exceeds OpenAI's 272,000-token threshold, so the whole request is billed at $2.00 for input and $7.50 for output, giving $0.83 per document instead of the $0.42 it would cost at base rates. That brings Terra level with Sonnet 5.5, which charges a flat $2.00 per million input tokens across its full one-million-token window, and with Argon's introductory price. Grok 4.7 doubles all its rates once a prompt reaches 200,000 tokens, which is why it costs about twice as much as Terra here, and Haiku 5.5 moves to its higher tier above 100,000 tokens but still finishes cheapest at $0.21 per document.

··········

HOW THE BENCHMARK CLAIMS LINE UP.

OpenAI's launch post compares Terra mainly with its own models and with one Anthropic model, and it presents most results as relative figures rather than absolute scores. The company says that on its Agents' Last Exam computer-use benchmark at maximum reasoning effort, Terra is more than three times as token-efficient as Claude Opus 5.5 at one quarter of its price per token, which OpenAI translates into roughly nine times lower cost per task. The price part of that statement can be checked: Opus 5.5 is listed by Anthropic at $4.00 input and $20.00 output, exactly four times Terra's rates. The token-efficiency part is OpenAI's own measurement.

OpenAI also says Terra comes within five points of Opus 5.5 on Terminal-Bench-Science 0.1 at maximum reasoning, at less than one fifth of the estimated cost, and that it solves 50.3% of Terminal-Bench 4.0 tasks at maximum effort against 21.5% for GPT-5.6 Terra. The company did not publish comparisons with Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 4 Argon, Gemini 3.8 Flash or Grok 4.7, and none of those providers has published results on Agents' Last Exam that would allow a cross-check.

The absence of shared, same-conditions results is the main limit of any comparison made on launch day. Each vendor reports its own selection of benchmarks, at its own reasoning settings, often with scaffolding that is not public. Readers comparing Terra with Sonnet 5.5 or Argon should treat price and context figures as verified and capability claims as vendor statements until independent evaluations appear.

··········

WHICH WORKLOADS EACH MODEL FAVOURS.

On the published numbers, and leaving aside the two budget models, Terra is the cheapest model in the group, and it pairs that price with a one-million-token window, a 128,000-token output limit and hosted computer use. Teams currently running Claude Sonnet 5.5 or Gemini 4 Argon for high-volume agentic work can test Terra at half the token price, and the gap with Argon widens to a factor of four on December 1, when Google's introductory pricing ends. The saving is largest for workloads that reuse long, stable context, because Terra's $0.05 cached-input rate is the lowest among the non-budget models.

Terra is less attractive for single requests that go beyond 272,000 tokens of input. In that range its effective price matches Sonnet 5.5, which keeps one flat rate across its full window, so the choice between the two for long-document analysis will depend on quality rather than price. Batch workloads benefit equally on OpenAI, Anthropic and Google, which all list a 50% batch discount, while xAI's 20% batch discount for Grok 4.7 is smaller.

Claude Haiku 5.5 and Gemini 3.8 Flash remain the price leaders. Haiku 5.5 is roughly ten times cheaper than Terra per token on short prompts and also supports Anthropic's computer toolset, so it is the obvious alternative for routing, extraction and subagent tasks where its capability is sufficient. Gemini 3.8 Flash is cheaper than Terra until the end of 2026, supports computer use in preview and accepts audio, video and PDF input, which Terra does not, but its scheduled January price increase puts it above Terra on both input and output.

··········

WHAT IS ESTABLISHED AND WHAT REMAINS UNCERTAIN.

The verified elements are the list prices, cache rates, batch discounts, long-prompt rules, context windows and output limits shown above, all taken from the providers' official documentation on October 26, 2026, together with the scheduled Google price changes on December 1, 2026 and January 1, 2027. On those figures, GPT-6.1 Terra costs half as much per token as Claude Sonnet 5.5 and introductory Gemini 4 Argon, undercuts Grok 4.7 substantially on cached context, and sits about ten times above Claude Haiku 5.5.

What cannot yet be established is cost per completed task across vendors. That figure depends on how many tokens each model actually uses, including reasoning tokens and tokenizer differences, and on success rates that only OpenAI has published for Terra, mostly in relative form. OpenAI's claims of nine times lower cost per task than Claude Opus 5.5 and three times faster completion than its own larger models are vendor results that independent testing has not yet confirmed, and no comparable figures exist against Sonnet 5.5, Argon or Grok 4.7. For buyers, the practical step is to run their own workloads through the official APIs and measure tokens per task, using the list prices above as the fixed part of the equation.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org