Gemini 4 Hydrogen vs Claude Sonnet 5.5 vs GPT-6.1 Terra
Compare Hydrogen, Sonnet 5.5 and GPT-6.1 Terra by coding, computer use, multimodal benchmarks and API task cost. See where each leads and current APIMaster access.
Published
Hydrogen is a strong candidate for everyday coding, multimodal knowledge work and agent execution; Sonnet 5.5 leads the cited terminal and computer-use rows; GPT-6.1 Terra has the lowest direct Standard token rates of these three. Google's launch table does not establish one winner for every workload. Choose using accepted task results, required tools and the price of the actual route you will buy.
Hydrogen is not yet available on APIMaster. Check current prices and availability on the model card. Discounted routes are rolling out.
| Model ID | Original price: direct Standard | APIMaster current input / output | Active public routes | Discount | Model card |
|---|---|---|---|---|---|
gemini-4-hydrogen |
$1.50 / $7.50 | No public route | 0 | Not available | Check Hydrogen |
claude-sonnet-5-5 |
$2 / $10 | From $0.4784832 / $2.392416 | 7 | Up to ~76.1% off | Sonnet routes |
gpt-6.1-terra |
$1 / $5, up to 272K input | From $0.1636877 / $0.8184385 | 2 | Up to ~83.6% off | Terra routes |
gemini-4-carbon |
$2 / $10 through 2026-11-30 | From $0.8341424 / $4.170712 | 3 | Up to ~58.3% off | Carbon routes |
Checked November 13, 2026, 01:13 UTC. USD per million input / output tokens, uncached non-Batch inference. Public enabled routes only; client-exclusive offers excluded. Sonnet ranges up to $2 / $10, Terra up to $1 / $5, Carbon up to $2 / $10. Input/output prices are paired by route; the live marketplace card controls current prices and availability. Official long-context, cache and service-tier conditions are separate.
For the model's release and API changes, see the Hydrogen overview. To compare providers for the same Hydrogen model, use the Cheap Hydrogen API guide.
GPT Trial(GPT-6 Astra, GPT-6 Sol, GPT-5.6, GPT-5.5, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-6.1-sol, gpt-6-astra, gpt-6-sol, and gpt-6-luna available)
Sign up and claim a $20 GPT trial
Coding benchmarks: Hydrogen wins SWE-sweep, Sonnet wins Terminal-Bench
Google's November launch table contains the following results. This is a reported-results comparison, not an APIMaster benchmark. The methodology combines Google's runs with provider reports and public leaderboards; it does not use one identical agent harness for all models.
| Benchmark | Hydrogen | Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweep | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1, cheating-adjusted | 57.0% | 63.9% | Not reported |
| Terminal-Bench-Science 0.1 | 51.4% | 59.9% | 58.3% |
Hydrogen leads SWE-sweep by 10.4 percentage points over Sonnet and 7.7 over Terra. Sonnet leads Hydrogen by 6.9 percentage points on Terminal-Bench 4.1 and 8.5 on the science variant. These are percentage-point gaps, not percentage improvements in all programming work. Terra's missing Terminal-Bench 4.1 result is not a zero.
Google ran Hydrogen SWE-sweep with Antigravity CLI 1.4.8 and safety filters enabled, reporting subtask-weighted reward. Terminal-Bench 4.1 used a mini-swe-agent vision harness at high effort, averaging five attempts per task. Competitor results there came from the official leaderboard. Keep Terminal-Bench 4.0 cost/latency charts separate from the 4.1 table.
For repository work, test your own issue fixes, regression suite and patch review. For terminal-heavy or science workflows, Sonnet's stronger cited rows justify including it in the evaluation even when Hydrogen wins other categories.
Knowledge work, multimodal files and computer use
| Benchmark / metric | Hydrogen | Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| Agents' Last Exam v1.1, rubric score | 71.6% | 60.4% | 69.8% |
| AutomationBench-AA | 76.1% | 71.8% | 60.2% |
| Vending-Bench 2, simulated bank balance | $14,549 | $11,165 | $9,107 |
| OSWorld-2.1, offline subset | 77.6% | 83.9% | 74.2% |
| MKA, Climb WSR accuracy | 53.4% | 52.0% | 43.8% |
| Chartography | 66.6% | 61.6% | 63.2% |
| HLE-Diamond | 58.8% | 53.3% | 47.3% |
Hydrogen leads the displayed knowledge-work, multimodal and reasoning rows. Sonnet leads the offline OSWorld computer-use subset. Vending-Bench's dollar figures are simulated benchmark balances, not API prices, customer revenue or returns you can expect.
The Google methodology PDF specifies a 122-task Agents' Last Exam v1.1 setup using the October 17 task version, before the October 26 update. OSWorld uses 82 offline tasks, with Hydrogen/Sonnet scores maximized over three single-attempt runs. Chartography, AutomationBench, MKA and HLE use other stated sources and setups. These rows have different units and aggregation; counting wins or averaging them hides those differences.
OpenAI's official Terra page emphasizes visual reasoning, precise computer use and writing. That makes Terra a relevant candidate for screen workflows, but a product description does not erase the mixed benchmark results above.
API price and the same-token task budget
Official uncached Standard rates are Hydrogen $1.50 / $7.50, Sonnet 5.5 $2 / $10, and Terra $1 / $5 per million input / output tokens. Hydrogen is 25% below Sonnet's rates but 50% above Terra's rates.
Assume 100,000 input and 20,000 billed output tokens per task, with 1,000 separate tasks so Terra remains within its short-context price tier. Include thinking in output. The calculations below hold token counts fixed; models may use different tokenizers and reasoning lengths in actual work.
| Model / purchase path | One task | 1,000 tasks | APIMaster vs same-model direct Standard |
|---|---|---|---|
| Hydrogen direct Google Standard | $0.30 | $300.00 | No APIMaster Hydrogen quote |
| Sonnet 5.5 direct Standard | $0.40 | $400.00 | Baseline |
| Sonnet 5.5 APIMaster from | $0.09570 | $95.70 | About 76.1% cheaper |
| Terra direct short-context Standard | $0.20 | $200.00 | Baseline |
| Terra APIMaster from | $0.03274 | $32.74 | About 83.6% cheaper |
The lowest advertised rate is not the same as the lowest cost per accepted task. A cheaper route that needs more retries or manual correction can cost more overall. This article has not measured acceptance rate, latency or route-level output quality.
Google Hydrogen Batch/Flex rates are $0.75 / $3.75; OpenAI Terra Batch/Flex rates are $0.50 / $2.50, and Anthropic Batch offers a 50% token discount. Direct cached input is cheaper again and needs separate cache-write/hit calculations. Terra prompts above 272K input tokens trigger whole-request 2x input/cache and 1.5x output rates; Hydrogen's current Standard rate is flat across its unified window. Tools, regions, taxes, platform fees and retries are excluded from the table.
Hydrogen vs Carbon and Argon: when to keep a frontier model
Google recommends Hydrogen for routine coding, responsive applications, multimodal documents and parallel execution. Carbon remains its strongest frontier model for hard architecture, science, security and difficult debugging. Argon targets advanced long-horizon work with lower token use than Carbon. This is a workload distinction, not a rule that Hydrogen must take every easy task or that Carbon wins every benchmark.
Google reported 51 Intelligence Index points for Hydrogen at default high, 54 at xhigh, and 50 for Argon at launch. Comparing Hydrogen xhigh with Argon default without stating settings would overstate a clean same-effort result. Hydrogen defaults to high; Carbon/Argon native defaults are medium.
Current Google Standard Hydrogen rates are 25% below Carbon/Argon's introductory $2/$10. Their scheduled December $4/$20 would make Hydrogen 62.5% cheaper, or 37.5% of their cost for fixed tokens. On APIMaster, however, today's Carbon/Argon starting quotes are already below direct Hydrogen Standard. That is a comparison of different models and purchase paths, not proof of Hydrogen pricing on APIMaster.
See Carbon vs Argon for their dedicated comparison and the Hydrogen overview for the Flash migration details.
How to run a fair comparison before migrating
Use the same saved inputs, tools, task limits and acceptance rubric. Declare reasoning settings explicitly and record total input, thinking/output, cache hits, retries, elapsed time and human corrections. Start with a small representative set rather than treating a launch leaderboard as a production guarantee.
Native and compatible endpoints have different feature contracts. Hydrogen's shared window is 1,048,576 tokens with 65,536 synchronous output tokens per hop; long decode requires continuation or background execution. Google's tools and thought signatures must be preserved correctly. Terra supports text/image input and text output; its official tool calling requires Responses. Verify a route's tool support before switching clients.
For current APIMaster access: register → create an API key → use https://apimaster.ai/v1 and gpt-6.1-terra or claude-sonnet-5-5 after checking the Terra card and Sonnet card. Hydrogen's model ID is gemini-4-hydrogen, but no public APIMaster route is listed at this check.
FAQ
Is Hydrogen better than Claude Sonnet 5.5?
It leads several rows in Google's reported table, including SWE-sweep and MKA. Sonnet leads Terminal-Bench 4.1, the science variant and the offline OSWorld subset. Select for your tasks rather than declaring a universal winner.
Is Hydrogen cheaper than GPT-6.1 Terra?
Not at direct uncached short-context Standard rates: Hydrogen is $1.50/$7.50 versus Terra's $1/$5. Context length, reasoning, cache and the purchased route can change real costs.
Did APIMaster run these benchmarks?
No. The numbers are transcribed from Google's official launch evaluation and its documented external sources. The budget table is arithmetic from published rates; it is not a paid performance test.
Should I replace Carbon or Argon with Hydrogen?
Test Hydrogen for recurring execution and retain Carbon or Argon where deeper reasoning improves accepted outcomes. Google supports mixed-model workflows; that does not establish identical behavior on every compatible endpoint.
Can I use all three through APIMaster now?
Sonnet 5.5 and Terra have public routes at the check. Hydrogen has none; discounted Hydrogen routes are being rolled out. Do not fund a Hydrogen migration until its actual route is listed.