Short answer: This is now mainly a comparison for teams maintaining existing integrations—not a recommendation for a new build. Anthropic retired Claude Sonnet 4 from its direct API on June 15, 2026. Google’s documentation is inconsistent: its Gemini API model page says Gemini 2.5 models remain available to prior active users, while the Google AI Studio model page says Gemini 2.5 Pro is deprecated, existing API access is limited until December 1, 2026, and requests after that date will route to Gemini 3.8 Flash and use that model’s rates. As of November 12, 2026, check the notice that applies to your exact project and serving channel before planning around Gemini 2.5 Pro. [1][2][3]
The performance findings below are historical, limited evaluations—not a current ranking. Vellum’s May 2025 article favored Sonnet 4 on its small SAT-math set, but named both Sonnet 4 and Gemini 2.5 Pro among its strongest modified-puzzle performers. It does not establish a universal winner. [7]
At a glance
| Factor | Claude Sonnet 4 — historical Anthropic API details | Gemini 2.5 Pro — Gemini API details |
|---|---|---|
| Identifier and status | claude-sonnet-4-20250514. Retired from Anthropic’s direct API on June 15, 2026; Anthropic says requests to retired models fail. Partner-operated services may have separate lifecycle schedules. [2] |
Google’s documentation differs by page: the Gemini API model page describes continued access for prior active users, while the AI Studio model page says access is limited until December 1, 2026, after which requests route to Gemini 3.8 Flash. Check the notice for your channel and project. [1][3] |
| Documented inputs and outputs | Anthropic’s system card describes visual-analysis capability for Sonnet 4. That alone does not establish complete API modality support or parity with Gemini. [8] | Google lists audio, image, video, text, and PDF inputs, with text output for the Developer API model. This is a positive statement of Gemini’s documented support, not proof that Sonnet lacked a corresponding capability. [3] |
| Context and output limits | Anthropic announced a one-million-token context option for Sonnet 4. The reviewed material does not establish all eligibility conditions or a model-specific maximum output limit. This historical announcement is not evidence of current API availability. [9] | Google lists a 1,048,576-token input limit and a 65,536-token output limit for the Developer API model. [3] |
| Price evidence | Anthropic’s launch announcement listed $3 per million input tokens and $15 per million output tokens. This is historical pricing for a model now retired from Anthropic’s direct API. [6] | A Google developer-forum post dated August 2025 quoted then-published rates of $1.25 per million input tokens and $10 per million output tokens for prompts up to 200,000 tokens; it says the output rate includes thinking tokens. I could not confirm that these remain the applicable rates on November 12, 2026. Treat them as historical context, not a current budget quote. [4][5] |
| Comparable latency | No comparable production latency measurement is established here. | No comparable production latency measurement is established here. |
Which model fits which workload?
- For a new Anthropic direct API integration: Sonnet 4 is not available through that channel. Check Anthropic’s current model lineup rather than assuming a newer model has the same capabilities or lifecycle. [2]
- For an existing Gemini 2.5 Pro integration: Verify access and the applicable retirement notice for the actual API channel. Google’s API reference and AI Studio model page give different lifecycle guidance, and the AI Studio page specifies a December 1, 2026 routing change. [1][3]
- If you need audio, video, or PDF inputs: Google lists these as Gemini 2.5 Pro Developer API inputs. That establishes Gemini’s documented support; the material reviewed here does not support treating them as proven differences from Sonnet 4 in every compatible channel. [3][8]
- For a new Gemini API project: Google recommends newer models in its Gemini API documentation. The AI Studio page also describes a December 1, 2026 transition for Gemini 2.5 Pro requests. [1][3]
What the benchmark comparison says
Vellum’s May 2025 article describes two limited evaluations:
- In 28 modified familiar puzzles, Vellum says Sonnet 4 answered 21 correctly and names both Sonnet 4 and Gemini 2.5 Pro among the strongest performers. The captured article does not give Gemini’s exact score.
- In 50 difficult SAT math questions, Vellum says Sonnet 4 scored highest and characterizes Gemini 2.5 Pro as around 50%, without giving an exact Gemini score in the captured text.
Vellum describes zero-shot prompts, an LLM judge, and human review, but the available article text does not provide enough detail to reproduce the results. These are narrow historical findings, not a current ranking for reasoning, math, coding, or production workloads generally. [7]
Cost and latency: what can be compared
The best Gemini rate detail retrieved for this comparison comes from an August 2025 Google AI Developers forum post, which quotes then-published Developer API rates of $1.25 per million input tokens and $10 per million output tokens, for prompts up to 200,000 tokens. The post says the output price includes thinking tokens. Google’s pricing page retrieved for this update did not expose a Gemini 2.5 Pro rate, so I could not confirm whether those historical figures remain applicable to existing users today. Check the live price and billing terms for your project before budgeting. [4][5]
For scale only, at those quoted historical rates, a request with 10,000 input tokens and 2,000 output tokens total (including any billed thinking tokens) would cost $0.0325: $0.0125 for input plus $0.02 for output. This is an illustration using the quoted rates, not a verified current Gemini quote. The forum excerpt does not establish rates for prompts over 200,000 tokens or other billing conditions.
Anthropic’s May 2025 launch announcement listed Sonnet 4 at $3 per million input tokens and $15 per million output tokens. At those historical launch rates, the same 10,000-input/2,000-output-token example would be $0.06. That is a historical calculation only; Sonnet 4 has been retired from Anthropic’s direct API and is not a current callable alternative there. [2][6]
Vellum’s 2025 article reports a 1.9-second latency figure but inconsistently attributes it to Sonnet 4 in one passage and Claude 3.7 Sonnet in another. It also supplies no test conditions. Do not treat that number as a confirmed Sonnet 4 measurement or compare it with Gemini. [7]
Validate an existing integration before deciding
- Confirm lifecycle and access for the exact serving channel and project. For Sonnet 4, Anthropic’s direct API retirement means you cannot newly test it there. For Gemini 2.5 Pro, check the applicable Google notice and any stated migration date. [1][2]
- Define representative inputs and pass criteria from your own workload—such as task success, acceptable errors, and required output format.
- Test only models you can actually call. Record each model ID, prompt, relevant thinking settings, and retry policy. Do not treat old benchmark results as a substitute for testing your available candidate.
- Compare quality, billed tokens, and latency on the same inputs and under the same retry conditions. Include thinking tokens in Gemini output-cost estimates where they are billed as output.
- Decide using your measured trade-offs and document the configuration and results so later model or lifecycle changes can be reassessed.
Bottom line
For new production work, Claude Sonnet 4 is no longer callable through Anthropic’s direct API. Gemini 2.5 Pro is a legacy option with differing Google lifecycle notices; one current AI Studio notice specifies a December 1, 2026 transition to Gemini 3.8 Flash, while the Gemini API model reference gives different availability wording. Historical Vellum results offer limited, task-specific evidence rather than a universal winner. Use the Gemini rate example above only as historical context until you confirm the rate that applies to your project, and validate existing integrations against models you can call through their actual serving channel. [1][2][3][7]
Sources
Gemini 2.5 Pro — Gemini API | Google AI Studio
Model deprecations — Claude Platform Docs
Billing Discrepancy for Gemini 2.5 Pro Usage — Google AI Developers Forum
Gemini Developer API pricing — Google AI Studio
Introducing Claude 4 — Anthropic
Evaluation: Claude 4 Sonnet vs OpenAI o4-mini vs Gemini 2.5 Pro — Vellum
Claude 4 System Card — Anthropic
Claude Sonnet 4 now supports 1M tokens of context — Anthropic