GPT vs Claude vs Gemini: API price and latency compared

In November 2026, GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon each list at $2 per million input tokens and $10 per million output tokens; in Requesty's streaming traffic, Sonnet 5.5 had the fastest median first token of the three at 2.8 seconds and the fastest output at 200 tokens per second. Sol measured 4.0 seconds and 71 tokens per second, while Argon measured 7.7 seconds and 84 tokens per second.

November 2026 API prices

These are standard prices per million tokens for short-context requests in November 2026, before cache discounts. GPT-6.1 Terra's standard price is $1 input and $5 output; the lower $0.50 and $2.50 rates belong to Terra Flex.

Sources: OpenAI pricing, Anthropic pricing, Gemini API pricing and Requesty's model list. For more models, see our LLM API price comparison.

Context length can change the bill. For the GPT-6 models in the table, requests above 272,000 input tokens have double the input rate and 1.5 times the output rate, applied to the full request.

Source: OpenAI pricing.

Google's output prices include thinking tokens (Gemini API pricing).

Measured latency depends on the model

Time to first token (TTFT) is the wait until a streaming response starts. Output speed is the rate at which tokens arrive after generation begins. A fast first token helps a response feel immediate; output speed also affects how long a longer answer takes.

The Requesty rankings below cover streaming responses from September 7 to November 9, 2026. P95 TTFT is the first-token wait at the 95th percentile, which gives a view of slower requests.

Two lighter Gemini models also had low first-token waits: Gemini 3.1 Flash-Lite measured 739 ms at 230 tokens per second, and Gemini 3.5 Flash-Lite measured 807 ms at 361 tokens per second.

These figures come from production traffic, with different cache usage across models. Argon accounted for 0.1% of requests in the rankings window.

First token is only part of the wait

Take a 500-token answer and use the rankings' median TTFT and output speed as inputs. For Sonnet 5.5, 2.786 seconds plus 500 / 200 seconds gives roughly 5.3 seconds. For Sol, 4.041 seconds plus 500 / 71 seconds gives roughly 11.1 seconds.

The estimate combines medians from different requests, so it is not a measured total latency.

Provider-level data makes the same distinction useful. In our April 2026 streaming note, Vertex hosting Gemini had a 1.29-second median TTFT and a 3.05-second median total latency. Those are older provider-level figures with their own traffic mix, separate from the model table above.

The April provider latency leaderboard also found that workload mix drove much of the gap between providers. Measure both first-token and total latency on the requests you intend to send.

Blended cost reflects the traffic

List price tells you the rate for input and output tokens. Blended cost in the Requesty rankings is what we billed per million tokens across observed traffic, with cache discounts applied.

Sol and Sonnet 5.5 share a $2 input and $10 output list price, but their observed blended costs were $0.43 and $1.14 per million tokens, respectively. That makes Sonnet around 2.65 times as expensive in this window. Their cache hit rates were 92.3% for Sol and 80.7% for Sonnet.

Cache rates differ by model. For example, Anthropic's pricing sets a Sonnet 5.5 cache hit at $0.10 per million input tokens, compared with $2 for standard input. Your own prompt repetition and input-to-output ratio determine how much of that discount you can use. Our prompt caching price comparison covers rates across providers.

A monthly price calculation

Take an application that sends 100,000 requests per month to Gemini 4 Argon, each with 1,000 input tokens and 200 output tokens. All tokens are billed at the standard rate, with no cache discount.

That is 100 million input tokens and 20 million output tokens per month. At $2 input and $10 output, the model cost is:

(100 x $2) + (20 x $10) = $400 per month

If those model costs were billed through Requesty at its 5% fee, the total would be $420. Cache discounts and a different token mix would change the bill.

Coding results change the choice

Price and speed leave out whether a model completes the job. Our coding benchmark, updated October 29, 2026, ran 129 private Go backend tasks with the same agent and an 80-turn budget.

Opus solved the highest share of tasks in this benchmark. Sol solved a smaller share than Sonnet at about half the cost per solved task ($0.30 versus $0.59). For a coding agent, that trade-off can be more useful than comparing token prices alone.

The benchmark measures speed end to end from one region, including queueing and network time. Its output speeds differ from the production rankings, and some providers do not count hidden reasoning in output tokens, so the benchmark's GPT throughput reads lower than the work done.

Which model should you try first?

Run the candidates on the same prompts and record task completion alongside median and P95 latency and cost. Requesty's observability dashboard separates TTFT from total latency and shows cost and error rates.

Requesty gives you one OpenAI-compatible API across models and providers, with latency-based routing and fallback chains. Before putting a route into production, compare the results on your own traffic and plan how it should behave when a provider fails.