Gemini 4 Hydrogen vs GPT-6.1 Terra and Claude Sonnet 5.5: API prices and task costs
Gemini 4 Hydrogen, the model Google released on November 12 as its fast "daily driver" for agentic work, costs $1.50 per million input tokens and $7.50 per million output tokens on the paid tier of the Gemini API, with implicit cache hits billed at $0.075. That places it between the two models developers are most likely to weigh it against: OpenAI's GPT-6.1 Terra, listed at $1.00 input and $5.00 output, and Anthropic's Claude Sonnet 5.5, listed at $2.00 input and $10.00 output. On short and medium prompts Hydrogen therefore costs exactly 1.5 times as much as Terra and 25% less than Sonnet 5.5, token for token.
The order changes once prompts get long. Google charges the same Hydrogen rate across the full 1,048,576-token window, and Anthropic does the same for Sonnet 5.5 across its 1M-token window, while OpenAI doubles Terra's input rate and raises its output rate by half for any request with more than 272,000 input tokens. In the long-document workload calculated below, a 600,000-token request costs $0.96 on Hydrogen, $1.26 on Terra and $1.28 on Sonnet 5.5, so Hydrogen becomes the cheapest of the three for that class of job.
All prices in the tables below are as listed on the providers' official pricing pages and model documentation on November 12, 2026, in USD per 1M tokens. Hydrogen's figures come from Google's launch post and the Google AI Studio documentation for the model, because the Gemini API pricing page had not yet added a Hydrogen row on that date. Google has not published a Batch price for Hydrogen.
··········
GEMINI 4 HYDROGEN, GPT-6.1 TERRA AND CLAUDE SONNET 5.5 LIST PRICES.
The three models occupy the middle of their vendors' lineups. Google positions Hydrogen below Gemini 4 Carbon and Gemini 4 Argon, which it describes as its frontier models for long-horizon reasoning, and above Gemini 3.5 Flash-Lite. OpenAI lists Terra below GPT-6.1 Sol and GPT-6 Astra and above GPT-6 Luna. Anthropic describes Sonnet 5.5 as the faster, lower-cost companion to Claude Opus 5.5 for well-scoped everyday tasks, bug fixing and office documents.
........
Item (USD per 1M tokens) | Gemini 4 Hydrogen | GPT-6.1 Terra | Claude Sonnet 5.5 |
Input | $1.50 | $1.00 | $2.00 |
Output (incl. reasoning) | $7.50 | $5.00 | $10.00 |
Cached input / cache read | $0.075 (implicit) | $0.05 | $0.10 |
Cache write | No fee listed | $1.25 (30 min), $2.50 (12 h) | $2.50 (5 min), $4.00 (1 h) |
Batch input / output | Not published | $0.50 / $2.50 | $1.00 / $5.00 |
Prompts over 272K tokens | Same rate | Input 2x, output 1.5x | Same rate |
Regional processing | Not stated | +10% | +10% (US-only inference) |
........
Two details shape the comparison more than the headline rates. Hydrogen supports only implicit caching: Google applies the $0.075 rate automatically to repeated prefixes in prompts of at least 4,096 tokens, and the model rejects explicit caching calls such as `cachedContents` and `caches.create`. OpenAI and Anthropic both charge to write a cache entry and then bill reads at 5% of the base input rate, the same 95% discount Google gives Hydrogen. Anthropic's minimum cacheable prompt for Sonnet 5.5 is 512 tokens, far below Hydrogen's 4,096-token threshold, which matters for chat applications with short system prompts.
The second detail is the free tier. Hydrogen is available free of charge, with rate limits, on the free tiers of Google AI Studio and the Gemini API. OpenAI's model page lists the free tier as not supported for Terra, with paid tiers starting at 5,000 requests per minute and 1,000,000 tokens per minute on the Build tier.
··········
CONTEXT WINDOWS, OUTPUT LIMITS AND LONG-PROMPT PRICING FOR THE THREE MODELS.
All three models advertise a context window of roughly one million tokens, but they divide it differently between input and output and price the upper part of it differently.
........
Specification | Gemini 4 Hydrogen | GPT-6.1 Terra | Claude Sonnet 5.5 |
Context window | 1,048,576 (shared) | 1,050,000 | 1M |
Max output per request | 65,536 per hop; up to 1M in background mode | 128,000 | 128K; 300K on Batch (beta) |
Input types | Text, images, PDFs, audio, video | Text, images | Text, images |
Reasoning controls | low, medium, high (default), xhigh | none to max, medium default | Adaptive, high default on API |
Sampling parameters | temperature, top_p, top_k ignored | Not stated | temperature, top_p, top_k return 400 |
Knowledge cutoff | Not stated | June 1, 2026 | June 2026 |
........
Hydrogen's output ceiling works differently from the other two. A synchronous call returns at most 65,536 tokens, the same per-hop limit Google documents for streaming and non-streaming requests. With `background: true` the server keeps generating in 64K hops until the shared 1,048,576-token budget, minus the input, is used up or the `max_output_tokens` value is reached. Without an explicit `max_output_tokens`, a background run stops at 65,536 tokens. Terra and Sonnet 5.5 return up to 128,000 tokens in a single response, and Anthropic allows 300,000 tokens on the Message Batches API with a beta header.
Input modalities are the clearest functional gap. Google lists audio and video, including agentic video understanding that inspects specific frames, among Hydrogen's inputs. OpenAI's model page lists text and image input for Terra and states that audio and video are not supported, and Anthropic lists text and images for Sonnet 5.5. A transcription-plus-analysis pipeline or a video review agent would need a separate model on the OpenAI and Anthropic side.
··········
WHAT A SUPPORT CHATBOT, A CODING AGENT AND LONG-DOCUMENT ANALYSIS COST ON EACH MODEL.
The workloads below use list prices only and the same token counts for every model. In practice the models will not produce identical token counts for the same task, because each spends a different number of reasoning tokens, and all three vendors bill reasoning tokens as output. Google says Hydrogen used 12% to 23% fewer tokens than Gemini 3.8 Flash in partner tests, and Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 because it uses fewer tokens, but no vendor has published a token-efficiency comparison against the other two. The figures are therefore a price comparison at equal token volume, and real bills will move with each model's verbosity.
The support chatbot handles 10,000 requests a day for 30 days. Each request carries 6,000 input tokens, of which 4,500 are a stable system prompt and knowledge-base prefix served from cache, plus 1,500 fresh tokens and 400 output tokens. Cache writes are modeled as one refresh every 30 minutes for Terra (48 a day) and one every 5 minutes for Sonnet 5.5 (288 a day), and Hydrogen's implicit caching carries no listed write fee. The coding agent session makes 40 model calls averaging 80,000 input tokens each, 90% of them cache hits, and 2,000 output tokens per call, repeated 500 times a month. The long-document job sends a 600,000-token file with no caching and receives 8,000 output tokens, repeated 1,000 times a month.
........
Workload | Gemini 4 Hydrogen | GPT-6.1 Terra | Claude Sonnet 5.5 |
Support chatbot, per day | $55.88 | $37.51 | $77.61 |
Support chatbot, per month | $1,676.25 | $1,125.28 | $2,328.31 |
Coding agent, per session | $1.30 | $0.86 | $1.73 |
Coding agent, 500 sessions | $648.00 | $432.00 | $864.00 |
600K-token document, per request | $0.96 | $1.26 | $1.28 |
600K-token document, 1,000 requests | $960.00 | $1,260.00 | $1,280.00 |
........
The chatbot arithmetic for Hydrogen is 1,500 fresh tokens at $1.50, 4,500 cached tokens at $0.075 and 400 output tokens at $7.50, which comes to $0.0055875 per request and $55.88 for 10,000 requests. The same request costs $0.003725 on Terra before cache writes and $0.00745 on Sonnet 5.5; the cache refreshes add about $0.26 a day on Terra and $3.11 a day on Sonnet 5.5. The coding session shows the same ratios because it stays well under Terra's 272,000-token threshold: each Hydrogen call costs 8,000 fresh tokens at $1.50, 72,000 cached tokens at $0.075 and 2,000 output tokens at $7.50, or $0.0324, and 40 calls add up to $1.30.
The long-document row reverses the order. On Terra, the 600,000-token prompt crosses the 272,000-token line, so the whole request is billed at $2.00 input and $7.50 output: $1.20 for the input and $0.06 for the output. Hydrogen bills the same request at its normal $1.50 and $7.50 rates, for $0.90 and $0.06. Sonnet 5.5 keeps its standard rates across the window, but its $2.00 base input price already equals Terra's surcharged rate, and its $10.00 output rate puts it at $1.28.
The results point to clear use cases. Terra is the lowest-cost option of the three for high-volume chat and agent loops with prompts under 272,000 tokens, by a constant one-third against Hydrogen. Hydrogen is the cheapest for single requests that load several hundred thousand tokens of documents, transcripts or code, and it is the only one of the three with native audio and video input. Sonnet 5.5 is the most expensive at equal token volume in every workload, so a team choosing it on cost grounds would need it to finish tasks in markedly fewer tokens or fewer retries. Teams that can wait for results can halve Terra's and Sonnet 5.5's costs with Batch processing, an option Google has not priced for Hydrogen.
··········
GEMINI 4 HYDROGEN AGAINST GEMINI 3.8 FLASH INTRODUCTORY PRICING UNTIL DECEMBER 31.
For existing Gemini API customers, the most immediate price comparison is inside Google's own lineup. Google's documentation for Gemini 3.8 Flash lists an introductory price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, with standard pricing of $1.50 input and $7.50 output taking effect on January 1, 2027. Hydrogen launched at that standard rate.
For the rest of 2026, moving a workload from Gemini 3.8 Flash to Hydrogen doubles the list price per token. The 600,000-token document job costs $0.48 per request on Gemini 3.8 Flash at the introductory rate, against $0.96 on Hydrogen. From January 1 the per-token rates are identical, and the choice depends on how many tokens each model spends on a task. Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking level, produces more than five times fewer fabricated claims, and used 12% to 23% fewer tokens in partner tests on graphs, SQL and data-heavy enterprise workflows. Those are vendor claims without independent replication so far.
Google's migration notes for teams moving from Gemini 3.8 Flash list several request changes: switch the model ID to `gemini-4-hydrogen`, remove low hardcoded `max_output_tokens` caps such as 8,192 or raise them to 65,536, set `thinking_level` explicitly because Hydrogen defaults to high where 3.8 Flash defaulted to medium, delete `cached_content` fields, and drop `temperature`, `top_p` and `top_k`, which the server ignores. Because Hydrogen also defaults to a higher thinking level than Carbon and Argon, a team that changes only the model ID may see more output tokens per request than it expects. Developers still running older Gemini models can check our report on how the Gemini API reroutes Gemini 2.5 Pro and 2.5 Flash from December 1.
··········
GEMINI 4 HYDROGEN AND CLAUDE SONNET 5.5 BENCHMARKS PUBLISHED BY GOOGLE AND ANTHROPIC.
Google and Anthropic each published benchmark results with their launches, but the two sets rarely measure the same thing under the same conditions. OpenAI's launch page for GPT-6.1 Terra could not be retrieved for this comparison, and its developer model page lists no benchmark figures, so Terra is left out of this table rather than represented by secondhand numbers.
........
Benchmark (vendor-reported) | Gemini 4 Hydrogen | Claude Sonnet 5.5 | Comparable |
Terminal-Bench | 57% (v4.1) | 70.6% (v4.0) | No, different versions |
Humanity's Last Exam | 58.8% (HLE-Diamond) | 64.5% (with tools) | No, different variants |
Artificial Analysis Intelligence Index | 51 (high), 54 (xhigh) | Not reported | No |
Agentic tax and legal research | Vals Tax Agent 52.3%, Vals Legal Research 55.8% | Not reported | No |
Office and agent tasks | Agents' Last Exam v1.1 71.6% | GDPval-AA v2.1 1844 Elo, OSWorld 2.1 80.1% | No |
........
The Terminal-Bench figures illustrate the problem. Google reports 57% for Hydrogen on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash, while Anthropic reports 70.6% for Sonnet 5.5 on Terminal-Bench 4.0. The versions differ, test setups and effort settings are not disclosed in a comparable form, and the higher Sonnet 5.5 number cannot be read as a lead of 13.6 points. The same applies to Humanity's Last Exam, where Google cites an HLE-Diamond subset and Anthropic a with-tools configuration.
The most useful cross-vendor reference Google gives is the Artificial Analysis Intelligence Index, an independent composite in which Hydrogen scores 51 at its default high setting, one point above Gemini 4 Argon, and 54 at the new xhigh setting, which spends more tokens per task and therefore costs more per answer at the same list price. Anthropic's Sonnet 5.5 announcement does not cite an index score, and none was available for Terra in the sources consulted.
··········
CACHING, BATCH AND API FEATURE DIFFERENCES DEVELOPERS HIT WHEN SWITCHING TO GEMINI 4 HYDROGEN.
Teams moving a Terra or Sonnet 5.5 workload to Hydrogen, or the reverse, will run into practical differences beyond the price list. Caching is the first. On Hydrogen there is nothing to configure and no write fee, but the discount only applies above 4,096 tokens and cannot be pinned with an explicit cache object, so cache behavior depends on Google detecting a repeated prefix. Sonnet 5.5 charges $2.50 per million tokens to write a five-minute entry, which pays off quickly at high request volume and can cost more than it saves on sporadic traffic. Terra's 30-minute and 12-hour cache retention options suit workloads with long gaps between requests.
Batch processing is the second. OpenAI and Anthropic both offer a 50% discount for asynchronous jobs, bringing Terra to $0.50 input and $2.50 output and Sonnet 5.5 to $1.00 and $5.00. Google's Hydrogen documentation points to the pricing page for batch rates, and on November 12 that page did not list Hydrogen, so overnight evaluation runs, bulk classification and document backfills cannot yet be priced on Hydrogen's Batch tier.
The third difference is API surface. Hydrogen works through both the GenerateContent API and Google's Interactions API, and Google recommends `previous_interaction_id` for multi-turn and tool-calling sessions, including sessions that route planning turns to Gemini 4 Carbon and execution turns to Hydrogen. Terra requires the Responses API for tool calling and supports hosted tools including web search, file search, code interpreter, computer use and MCP. Sonnet 5.5 is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, which gives it the widest choice of cloud billing among the three. Google's Hydrogen announcement does not mention Vertex AI availability.
For subscription access and the full Gemini model price list, see our guide to Gemini API pricing and plans. The launch details for Hydrogen in the Gemini app and Google Search are covered in our report on the Gemini 4 Hydrogen launch, and an earlier comparison of GPT-6.1 Terra with Claude Sonnet 5.5 and Gemini 4 Argon covers computer use and the OpenAI and Anthropic models in more depth.
··········
GEMINI 4 HYDROGEN PRICING VERSUS TERRA AND SONNET 5.5: CONFIRMED FIGURES AND OPEN QUESTIONS.
The confirmed numbers settle the price ranking for most workloads. At equal token volume and prompts under 272,000 tokens, Hydrogen costs 50% more than GPT-6.1 Terra and 25% less than Claude Sonnet 5.5 for input and output alike, and its cache-read price sits between the two. Above 272,000 input tokens, OpenAI's surcharge lifts Terra's input rate to $2.00, and Hydrogen becomes the lowest-priced of the three for single long-context requests, with no change in its rate up to the full window.
Several figures that would change real-world costs are still missing. Google has not published Batch, Flex or Priority prices for Hydrogen, nor its paid-tier rate limits, and the main Gemini API pricing page had not been updated for the model on launch day. No vendor has published per-task token counts that allow a like-for-like efficiency comparison across Hydrogen, Terra and Sonnet 5.5, and the benchmark results available use different versions and configurations. Until those numbers exist, the list-price arithmetic favors Terra for high-volume short-prompt traffic, Hydrogen for long-context and audio or video inputs, and leaves Sonnet 5.5 to justify its higher price through task completion rates that buyers will need to measure on their own workloads.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
