# Claude Sonnet 5.5 vs Gemini 3.8 Flash: Cost per Task, Tested

> Gemini 3.8 Flash is 2.67x cheaper per token, yet on 11 identical calls at high effort on October 26, 2026 it cost 2.42x as much as Claude Sonnet 5.5.

- URL: https://blog.laozhang.ai/en/posts/claude-sonnet-5-5-vs-gemini-3-8-flash
- Published: 2026-10-26
- Updated: 2026-10-26
- Author: LaoZhang AI Team (https://blog.laozhang.ai/en/about)
- Topic: Model Comparisons
- Tags: Claude Sonnet 5.5, Gemini 3.8 Flash, Model Comparison, API Pricing, Claude API, Gemini API

---
For API work that needs reasoning, such as coding, parsing with edge cases or multi-step math, make Claude Sonnet 5.5 (`claude-sonnet-5-5`) your default, even though Gemini 3.8 Flash (`gemini-3.8-flash`) lists at 2.67 times less per token. In a same-prompt test on October 26, 2026, both models at high effort passed all 11 calls, but Flash wrote 80,277 output tokens to Sonnet's 9,383, almost all of it hidden thinking. So Flash cost $0.310 to Sonnet's $0.128 and took 702 seconds to Sonnet's 86.

Keep Gemini 3.8 Flash for jobs where a long input produces a short answer, like summarizing a document, and for video or audio input, which Sonnet doesn't accept. On a 6,723-word summary Flash cost $0.016 against Sonnet's $0.033. That edge has a date on it: Flash's price doubles on January 1, 2027, which puts the summary job at about break-even and the whole test at 4.84x Sonnet's cost.

The per-token figure you saw on a price page is accurate, but you pay per finished task, and Flash at high thinks far longer to reach the same answer. Artificial Analysis's harder benchmark tasks show the same direction: at an equal score, Flash costs 2.6x as much per task as Sonnet.

## Same prompts on both models, October 26, 2026: Flash cost 2.42x as much

On these eight tasks, Flash got every answer right and still cost 2.42 times as much as Sonnet 5.5, because it needed 8.6 times as many output tokens. The test ran on October 26, 2026, between 11:41 and 11:55 China time, through the laozhang.ai API gateway with one key for both models, so the route was the same. Sonnet 5.5 went through the Anthropic Messages endpoint with `output_config.effort: "high"`, which is the Claude API default, and no `thinking` field, so it used adaptive thinking with the thinking text left out of the response. Flash went through Gemini's native `generateContent` endpoint with `thinkingConfig.thinkingLevel: "high"`. Both had a 32,000-token output cap, no system prompt and no tools.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

"High" is the closest like-for-like setting, not an identical one. It is Flash's top thinking level, while Sonnet has `xhigh` and `max` above it. Five easier tasks ran once each: a word problem, three small Python functions graded against hidden tests, and a five-bullet summary of a 6,723-word article. Three harder tasks ran twice each: an arithmetic expression evaluator written without `eval`, a CSV line parser written without the `csv` module, and a count of the numbers below 1,000,000 whose digits sum to 27. All 22 answers passed. Every function passed all of its hidden tests, both models found 55,252 on both runs, and both summaries had exactly five bullets under 25 words. On these tasks quality tied, so the table shows what each correct answer cost.

Costs are the returned token counts times the Standard list prices. Flash reports its output as thinking plus answer. Sonnet reports a single output count that already includes its hidden thinking. For the three harder tasks the token counts are per run and the cost and seconds are the mean of two runs. The total row sums all 11 calls.

| Task (hidden tests) | Sonnet 5.5 tokens in / out | Sonnet cost | Sonnet time | Flash tokens in / thinking + answer | Flash cost | Flash time | Flash ÷ Sonnet cost |
|---|---|---|---|---|---|---|---|
| Word problem | 86 / 704 | $0.0072 | 7.1 s | 61 / 1,716 + 355 | $0.0078 | 78.5 s | 1.08x |
| `merge_intervals` (8) | 105 / 1,217 | $0.0124 | 9.3 s | 61 / 974 + 522 | $0.0057 | 25.6 s | 0.46x |
| `parse_duration` (18) | 248 / 723 | $0.0077 | 7.5 s | 207 / 4,463 + 318 | $0.0181 | 30.5 s | 2.34x |
| `top_words` (10) | 174 / 160 | $0.0019 | 3.0 s | 115 / 1,115 + 130 | $0.0047 | 27.2 s | 2.44x |
| 5-bullet summary of a 6,723-word article | 15,335 / 277 | $0.0334 | 5.0 s | 11,198 / 1,733 + 154 | $0.0155 | 14.3 s | 0.46x |
| `eval_expr` without `eval` (25), 2 runs | 314 / 1,816 and 1,849 | $0.0190 | 13.7 s | 212 / 14,876 + 730 and 30,721 + 809 | $0.0885 | 156.9 s | 4.67x |
| `parse_csv_line` without `csv` (16), 2 runs | 244 / 880 and 432 | $0.0070 | 6.1 s | 159 / 4,175 + 371 and 6,167 + 371 | $0.0209 | 34.2 s | 2.97x |
| Digit-sum count (answer 55,252), 2 runs | 69 / 678 and 647 | $0.0068 | 7.0 s | 51–52 / 4,831 + 847 and 4,034 + 865 | $0.0199 | 71.9 s | 2.94x |
| **Total, 11 calls** | 17,202 / 9,383 | **$0.128** | 86 s | 12,487 / 74,805 + 5,472 | **$0.310** | 702 s | **2.42x** |

Flash was cheaper on only two tasks, `merge_intervals` and the long summary, at 0.46x each, and about even on the word problem. On the other five it cost 2.3x to 4.7x as much. Split by difficulty, the five easier tasks cost $0.0627 on Sonnet and $0.0518 on Flash, so Flash was 17% cheaper there. The three harder tasks cost $0.0655 on Sonnet and $0.2586 on Flash, 3.95 times as much. The harder the problem, the longer Flash thought. Its second `eval_expr` run used 31,530 of the 32,000-token cap, twice the first run's 15,606, so a lower cap could have cut that answer off. Per-task figures this noisy are a reason to trust the totals more than any single row.

![Bar chart of Gemini 3.8 Flash's cost relative to Claude Sonnet 5.5 on each test task, from 0.46x to 4.67x, next to output tokens showing 74,805 of Flash's 80,277 were thinking](https://blog.laozhang.ai/posts/en/claude-sonnet-5-5-vs-gemini-3-8-flash/img/same-prompt-cost-per-task.webp)

What wasn't tested matters as much. This is one gateway on one day, with one or two runs per task, all at high. There are no medium or low runs, no agent loops, tool calls, multi-turn chats, caching or Batch jobs, no direct Anthropic or Google endpoints, and no text other than English. Both models solved every one of these small, well-specified tasks, so the test says nothing about where Flash fails. For that, use the benchmark scores in the next sections, and re-run a sample of your own prompts before you switch a production default.

## Why a model 2.67x cheaper per token costs more per task

Per-token price only decides the bill if both models write about the same number of tokens, and at high effort they don't. Both vendors bill thinking tokens as output. Flash spent 74,805 tokens thinking across the 11 calls, which is 93% of its output and eight times Sonnet's entire output, thinking included. Flash's output was 8.56 times Sonnet's while its output price is 2.67 times lower, so its output bill alone came out about 3.2 times higher.

You can recompute any row of the table with one formula, using Standard prices per million tokens. Sonnet 5.5 charges $2 input and $10 output. Gemini 3.8 Flash charges $0.75 and $3.75 through December 31, 2026. Cost is input tokens × input price + output tokens × output price, divided by 1,000,000, with thinking counted inside output. For the long summary that gives 15,335 × $2 + 277 × $10 = $0.0334 on Sonnet and 11,198 × $0.75 + 1,887 × $3.75 = $0.0155 on Flash. Flash wins this row because the bill is mostly input and the answer is short, so the thinking overhead has little to multiply.

The two models also count the same text differently. Sonnet 5.5 reported 15,335 input tokens for the 6,723-word article, while Flash reported 11,198 for the identical prompt. Anthropic's count was 37% higher, though it includes Anthropic's message wrapper. On this English text the real input-price gap was about 2.67 × 1.37 ≈ 3.65x rather than 2.67x, which is the other reason long-input jobs lean toward Flash. That's one text in one language. Code or other languages may tokenize differently.

Artificial Analysis (AA) measures cost per task on a much harder, multi-domain set, and as of October 26, 2026 it points the same way. Compared at an equal Intelligence Index score of 41, Sonnet 5.5 at medium costs $0.48 and 121 seconds per task, while Flash at high costs $1.24 and 498 seconds. That makes Flash 2.6x the cost and 4.1x the time for the same score in AA's harness. With both at high, Sonnet scores 47 to Flash's 41 and costs $0.88 per task against $1.24, which is 29% less. Flash at its default medium level scores 40 for $0.93, still more per task than Sonnet at high. AA's own headline, "For cost per task, Claude Sonnet 5.5 is cheapest", compares Sonnet at low ($0.35) with Flash at medium ($0.93). That's two different settings, so the equal-score pair above is the fairer comparison. All figures are from the [Artificial Analysis Sonnet 5.5 vs Gemini 3.8 Flash comparison page](https://artificialanalysis.ai/models/releases/comparisons/claude-sonnet-5-5-vs-gemini-3-8-flash).

## Is Gemini Flash better than Claude Sonnet? 41 vs 47 on AA's index

No, not overall. With both at high effort, Gemini 3.8 Flash scores 41 on Artificial Analysis's Intelligence Index v4.3.2 to Claude Sonnet 5.5's 47, as of October 26, 2026, and Sonnet reaches 56 at `max`, a level Flash doesn't have. Flash is ahead on a few individual evaluations, though, and they map to real jobs: factual recall, reading long documents, and by a hair on scientific coding and expert Q&A. Sonnet's lead is widest on terminal-based agentic coding and on office-style deliverables, where Flash at high trails by a large margin.

| AA evaluation (both at high) | Sonnet 5.5 | Gemini 3.8 Flash | What it suggests |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic terminal coding) | 43.9% | 19.7% | Coding agents: Sonnet, by a wide margin |
| AA-Briefcase | 1635 | 1202 | Business documents and deliverables: Sonnet |
| GDPval-AA | 1548 | 1436 | Professional work tasks: Sonnet |
| CritPt | 24.6% | 18.3% | Hard physics reasoning: Sonnet |
| AutomationBench-AA | 59.4% | 59.9% | Tool-driven automation: tie |
| SciCode | 53.7% | 56.6% | Scientific coding: slight Flash edge |
| Humanity's Last Exam | 45.8% | 47.8% | Expert Q&A: slight Flash edge |
| AA-LCR v1.1 (long-context reasoning) | 78.0% | 81.3% | Reading long documents: Flash |
| AA-Omniscience | 21 | 30 | Factual recall with fewer confident errors: Flash |

The gap on coding agents survives even when the two models are matched on index score. Sonnet at medium, which scores the same 41 as Flash at high, still gets 29.8% on Terminal-Bench to Flash's 19.7%, while Flash keeps its long-context (81.3% vs 76.3%) and Omniscience (30 vs 20) leads. Flash at its default medium level does even better on long-context reasoning, 84.0%, the highest of the eight settings AA lists for the two models. Keep these AA numbers apart from vendor-reported ones: Anthropic's own table lists Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, which comes from a different harness and settings and can't be set against AA's 19.7% for Flash. Single-user reports follow the AA split. In a linux.do thread from October 11, 2026, one user said 3.8 Flash at high, used through a Gemini subscription, stopped about 60% of the way through a dynamic-chart task after three or four retries, while Sonnet 5.5 on the Claude Free plan finished it. Another reply in the same thread rated Flash ahead for natural-sounding writing.

Taken alone, 3.8 Flash is a solid model, and Google calls it its most intelligent Flash model, built for long-horizon software work and agents. Whether to stay on it or move up within Google is covered in [Gemini 4 Argon vs 3.8 Flash: Cost per Task and When to Switch](https://blog.laozhang.ai/en/posts/gemini-4-argon-vs-gemini-3-8-flash).

## Which is faster? Sonnet 5.5 returns answers sooner at high effort

If "faster" means how soon you have a finished answer, Sonnet 5.5 at high was faster on every one of the 11 calls in the October 26 test, 86 seconds in total against Flash's 702. Speed pages disagree because they measure different things. [OpenRouter's Sonnet 5.5 vs Gemini 3.8 Flash comparison](https://openrouter.ai/compare/anthropic/claude-sonnet-5.5/google/gemini-3.8-flash), with data updated October 5, 2026, reports a median throughput of 108.0 tokens per second for Sonnet and 52.0 for Flash, based on real traffic at whatever settings callers used. AA, testing at fixed settings, has Flash streaming faster at high (137 tokens per second against Sonnet's 106).

Both are true, and neither tells you when the answer arrives. Tokens per second only measures speed once text is flowing. At high, Flash thinks for a long time before its first visible token and then writes far more tokens in total. AA records a time to first token of 27.26 seconds for Flash at high against 10.46 seconds for Sonnet at high and 1.35 seconds at medium, and a time per task of 498 seconds against 217. A model can stream faster and still finish later. The test's seconds are wall-clock time for single calls through one gateway, not a speed benchmark, but every call pointed the same way: the word problem took 78.5 seconds on Flash and 7.1 on Sonnet.

Flash at medium or low was not measured here, and AA publishes no speed figures for those levels. If Flash feels slow inside Antigravity specifically, [Gemini 3.8 Flash Slow in Antigravity? Find the Delay Before Retrying](https://blog.laozhang.ai/en/posts/gemini-3-8-flash-antigravity-slow) separates the model's thinking time from other delays.

## What changes on January 1, 2027, when Flash's price doubles

From January 1, 2027, every Gemini 3.8 Flash cost doubles, so most of the jobs where Flash still beats Sonnet stop being cheaper. Google's [Gemini API pricing page](https://ai.google.dev/gemini-api/docs/pricing) shows every 3.8 Flash rate doubling on that date: Standard input and output, Batch, Flex, Priority and context caching. As of October 26, 2026, Anthropic's [Claude API pricing page](https://platform.claude.com/docs/en/about-claude/pricing) lists no change for Sonnet 5.5. The table recomputes the test and AA's figures, assuming each model would use the same tokens after the change.

| Same work, same token counts | Sonnet 5.5 | Flash through Dec 31, 2026 | Flash from Jan 1, 2027 |
|---|---|---|---|
| Standard price per 1M tokens, input / output | $2 / $10 | $0.75 / $3.75 | $1.50 / $7.50 |
| All 11 test calls | $0.128 | $0.310 | $0.621 (4.84x Sonnet) |
| Five easier test tasks | $0.0627 | $0.0518 | $0.1036 (1.65x Sonnet) |
| 10,000 summary calls shaped like the 6,723-word task | $334 | $155 | $310 |
| 10,000 calls shaped like the `eval_expr` task | $190 | $885 | about $1,770 |
| AA cost per task, both at high | $0.88 | $1.24 | about $2.48 |

The 10,000-call rows multiply the measured per-call costs and are an illustration, not a measurement. Your token counts will differ, and Flash's `eval_expr` runs alone varied by a factor of two. The pattern is still clear. After the change, the two tasks where Flash won, the long summary and `merge_intervals`, fall to 0.93x and 0.91x of Sonnet's cost, and Flash at high costs about 5.2 times as much as Sonnet at medium per AA task at the same score. If you picked Flash for a pipeline on price alone, re-run a sample in December before the new rates apply.

## Which jobs to send to Sonnet 5.5 and which to Gemini 3.8 Flash

Default to Sonnet 5.5 for coding, agents and anything with a hard reasoning step, and route long-input reading, video and audio, and free prototyping to Flash. The table pairs each job with the evidence behind it and the condition that would flip it.

![Job routing map: coding agents, parsing and business documents go to Claude Sonnet 5.5; long documents, video or audio and free prototyping go to Gemini 3.8 Flash; automations and factual Q&A need testing on both](https://blog.laozhang.ai/posts/en/claude-sonnet-5-5-vs-gemini-3-8-flash/img/which-jobs-sonnet-or-flash.webp)

| Job | Default | Why | Route to the other model when |
|---|---|---|---|
| Coding agents, terminal work, bug fixing | Sonnet 5.5 | Widest AA gap (Terminal-Bench 4.0); the hardest test task cost 4.67x more on Flash | You work in Antigravity on Free or Plus, where Sonnet 5.5 isn't offered |
| Short reasoning calls: parsing, extraction with edge cases, math | Sonnet 5.5 | Flash cost 2.3x–3x more on `parse_duration`, `top_words`, the CSV parser and the digit count, and took 4–10x longer | The call is simple enough for Flash at low, which wasn't tested here |
| Long document in, short answer out | Gemini 3.8 Flash through Dec 31, 2026 | Half of Sonnet's cost on the summary; Flash leads on AA's long-context test | From January 1, 2027, the cost is roughly even, so test both again |
| Video or audio input | Gemini 3.8 Flash | Sonnet 5.5 takes text and images only | Only the text and images matter |
| Factual Q&A without retrieval | Test both; Flash has the edge | Flash leads on AA-Omniscience | You add retrieval and the gap closes on your own checks |
| Business documents, reports, office deliverables | Sonnet 5.5 | Clear leads on AA-Briefcase and GDPval-AA | Volume is huge and every draft gets human review anyway |
| Tool-driven automations | Either; decide on cost per run | AutomationBench-AA is a tie | Your own runs show one model using far fewer tokens |
| Prototyping with no budget | Gemini 3.8 Flash free tier | Free input and output on the Gemini API | Your prompts must not be used to improve Google's products |

For coding agents, set Sonnet's effort deliberately. Anthropic's effort guidance suggests `high` as the starting point for most API work, `medium` for well-specified agentic coding, and `medium` or `low` for chat. Claude Code already defaults Sonnet 5.5 to `medium`. Lower effort is also where AA's per-task cost for Sonnet drops to $0.48 at medium and $0.35 at low.

If neither model fits, the next step is usually a different tier rather than the other model. To cut cost on simple calls, see [Claude Haiku 5.5 vs Sonnet 5.5: When the 20x Price Gap Pays Off](https://blog.laozhang.ai/en/posts/claude-haiku-5-5-vs-sonnet-5-5). For harder agent work, see [Claude Sonnet 5.5 vs Opus 5.5: When Half the Price Really Saves](https://blog.laozhang.ai/en/posts/claude-sonnet-5-5-vs-opus-5-5). OpenAI's mid-tier is compared in [Claude Sonnet 5.5 vs GPT-6 Astra: When the Cheaper Model Wins](https://blog.laozhang.ai/en/posts/claude-sonnet-5-5-vs-gpt-6-astra), and a cross-vendor price table by input and output mix is in [AI API Price Comparison: Cheapest Model by Input and Output Mix](https://blog.laozhang.ai/en/posts/cheapest-llm-models).

## Specs side by side: context, output limit, inputs, caching, batch

The two models have the same 1M-token context class but differ in output length, input types and how you control reasoning. Those differences decide some jobs before price does. Figures are from Anthropic's model overview and pricing pages and Google's Gemini 3.8 Flash model and pricing pages, as of October 26, 2026.

| Spec | Claude Sonnet 5.5 | Gemini 3.8 Flash |
|---|---|---|
| API model ID | `claude-sonnet-5-5` | `gemini-3.8-flash` |
| Released | September 28, 2026 | Generally available September 2, 2026 |
| Context window | 1M tokens, same per-token rate across the whole window | 1,048,576 tokens, no long-prompt surcharge |
| Max output | 128K tokens (300K on Batches with a beta header) | 65,536 tokens |
| Input types | Text, images, PDF ([document blocks](https://platform.claude.com/docs/en/build-with-claude/pdf-support), up to 600 pages / 32 MB per request) | Text, images, video, audio, PDF |
| Reasoning control | `effort`: `low`, `medium`, `high`, `xhigh`, `max`; adaptive thinking can't be disabled | `thinkingLevel`: `low`, `medium`, `high`; `minimal` returns an error |
| Default reasoning setting | `high` on the API | `medium` |
| Batch price per 1M tokens, input / output | $1 / $5 | $0.375 / $1.875, then $0.75 / $3.75 from 2027 |
| Cache reads per 1M tokens | $0.10 | $0.075, then $0.15 from 2027 |
| Cache writes and storage per 1M tokens | $2.50 (5-minute) or $4 (1-hour) to write | $0.50 per hour of storage, then $1.00 from 2027 |
| Free API use | A small amount of free credits for new users | Free tier: input, output and caching free of charge |
| Prompts used to improve products | Not covered on the pricing page | Yes on the free tier, no on paid |

Two rows change the math for repeated long prompts. Sonnet 5.5's cache reads fell to $0.10 on October 7, 2026. From 2027 that is cheaper than Flash's $0.15, though Flash charges storage by the hour instead of a write fee. Caching wasn't part of the test. [Claude Sonnet 5.5 Cache Read Price Cut to $0.10: What You Save](https://blog.laozhang.ai/en/posts/claude-sonnet-5-5-cache-read-price) works through the savings, and the full Gemini rate card is in [Gemini API Pricing 2026: Token Rates, Free Tier, and Key Cost](https://blog.laozhang.ai/en/posts/gemini-api-pricing). Google publishes no request-per-day table for the free tier. [Gemini API Free Tier 2026: Limits, Free Models, and Free API Keys](https://blog.laozhang.ai/en/posts/gemini-api-free-tier) covers what is known.

## Sonnet 5.5 and 3.8 Flash in Claude.ai, Gemini, Claude Code and Antigravity

Outside the API, Sonnet 5.5 is the easier of the two to reach for free, and 3.8 Flash is the easier one inside Google's Antigravity coding tool. Anthropic says anyone can chat with Sonnet 5.5 on Claude.ai, and as of October 26, 2026 the Claude apps offer an effort selector from Low to Max for it, with thinking that can't be turned off. Anthropic's help center doesn't say which effort level the Free plan gets. Claude Code runs Sonnet 5.5 at `medium` effort by default. Whether a Claude plan or an API key is the better way to pay for it is covered in [Claude Pro vs Max: Usage Limits, Fable, and When Max Is Worth It](https://blog.laozhang.ai/en/posts/claude-code-pro-vs-max).

On Google's side, the Gemini app's help page lists a model called "Flash", with no version number, on Google AI Plus, Pro and Ultra. With no plan you get Auto or Flash-Lite, and Auto mostly answers with Flash-Lite but may hand harder prompts to Flash. Antigravity's model table, as of October 23, 2026, offers 3.8 Flash on every plan from Free & Plus up to Enterprise. Sonnet 5.5 (thinking) is not available on Free & Plus there. It appears only on paid Google AI Pro (not trials), Ultra and Enterprise, so a free Antigravity user comparing the two can only run Flash.

## Claude Sonnet 5.5 vs Gemini 3.8 Flash: other questions

### Is Claude Sonnet 5 the same as Sonnet 5.5?

No. Claude Sonnet 5 is the older model, and Sonnet 5.5, released September 28, 2026, is its successor. They share base token prices, but Sonnet 5's cache reads still cost $0.20 per million tokens. Some comparison sites, BenchLM and Requesty among them in October 2026, list `claude-sonnet-5` against Gemini 3.8 Flash, so check the version before you trust a score or a cost figure.

### Can I turn thinking off to make either model cheaper?

Not fully on either. On the Claude API, Sonnet 5.5 rejects `thinking: {"type": "disabled"}` and fixed `budget_tokens` with a 400 error. You can lower `effort`, or send `thinking: {"type": "between_tools"}` to skip up-front thinking at `high` or below, according to [Anthropic's extended thinking documentation](https://platform.claude.com/docs/en/build-with-claude/thinking) as of October 26, 2026. Gemini 3.8 Flash's lowest setting is `thinkingLevel: "low"`, and `minimal` returns an error. Lower levels weren't part of the October 26 test, so measure your own prompts before you assume the savings.

### What happened to Gemini 3.7 Flash?

Google deprecated `gemini-3.7-flash` on October 8, 2026, and calls to it are now routed automatically to `gemini-3.8-flash`. If you still send the 3.7 ID, you are already being billed for, and getting answers from, 3.8 Flash.

## Sources

External pages this guide links to, in the order they appear. Last updated 2026-10-26.

- [Artificial Analysis Sonnet 5.5 vs Gemini 3.8 Flash comparison page](https://artificialanalysis.ai/models/releases/comparisons/claude-sonnet-5-5-vs-gemini-3-8-flash) (artificialanalysis.ai)
- [OpenRouter's Sonnet 5.5 vs Gemini 3.8 Flash comparison](https://openrouter.ai/compare/anthropic/claude-sonnet-5.5/google/gemini-3.8-flash) (openrouter.ai)
- [Gemini API pricing page](https://ai.google.dev/gemini-api/docs/pricing) (ai.google.dev)
- [Claude API pricing page](https://platform.claude.com/docs/en/about-claude/pricing) (platform.claude.com)
- [document blocks](https://platform.claude.com/docs/en/build-with-claude/pdf-support) (platform.claude.com)
- [Anthropic's extended thinking documentation](https://platform.claude.com/docs/en/build-with-claude/thinking) (platform.claude.com)