Table of Contents
Gemini 3.6 Flash vs Claude Sonnet 5 vs GPT-5.6 Terra: Pricing and Which to Use (July 2026)
Gemini 3.6 Flash landed on July 21, 2026 at $1.50 per million input tokens and $7.50 per million output tokens, per Google's Gemini API pricing. Claude Sonnet 5 runs $2/$10 through August 31, then steps to $3/$15, per TechCrunch. GPT-5.6 Terra is $2.50/$15, per OpenRouter. Run a workflow that generates 5 million output tokens a month and the bill is $37.50 on Gemini 3.6 Flash, $50 on Sonnet 5 at intro pricing, and $75 on Terra. Same job, double the cost at the wrong end of that spread.
So here is the pick, up front. For high-volume, cost-sensitive automation where a small quality gap is acceptable, Gemini 3.6 Flash is the one to route to. For draft generation and reasoning tasks where quality is worth paying for, Sonnet 5 while the intro rate holds. For anything that needs a million-token context window, GPT-5.6 Terra. For the hardest multi-step coding and reasoning, none of these three: that is Fable 5 and GPT-5.6 Sol territory, which we cover in the Fable 5 vs Sonnet 5 pricing guide. The rest of this walks through why, with the numbers under each call.
Three models, three price points
All figures below are per million tokens (MTok), input and output, taken from each provider's own pricing on July 22, 2026.
| Model | Provider | Input ($/MTok) | Output ($/MTok) | Reasoning control | Source |
|---|---|---|---|---|---|
| Gemini 3.6 Flash | $1.50 | $7.50 | Configurable (thinking budget) | ||
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | Configurable | ||
| Claude Sonnet 5 (intro, to Aug 31) | Anthropic | $2.00 | $10.00 | No public dial | TechCrunch |
| Claude Sonnet 5 (from Sep 1) | Anthropic | $3.00 | $15.00 | No public dial | TechCrunch |
| GPT-5.6 Terra | OpenAI | $2.50 | $15.00 | Reasoning param | OpenRouter |
| GPT-5.6 Luna | OpenAI | $1.00 | $6.00 | Reasoning param | Finout |
The Flash tier has two ways to get cheaper still. Batch and flex both run $0.75/$3.75, half the standard rate, and cached input drops to $0.15/MTok, per Google's pricing page. Terra caches input at $0.25/MTok, per the same OpenRouter listing. If a workflow re-sends the same system prompt or document on every call, caching is where the real savings hide, and all three providers price it separately from the sticker.
I put Luna in the table because it changes the OpenAI read. At $1/$6, Luna undercuts Gemini 3.6 Flash on output. Terra is the tier you reach for when you need its context window, not its price, and Luna is the value option in that lineup, per Finout's tier breakdown.
The cost math at volume
Output tokens dominate an LLM bill for most automation, because you send a short prompt and get back paragraphs. Here is the output-only spend at three volumes, at each model's list rate above.
| Monthly output tokens | Gemini 3.6 Flash | Sonnet 5 (intro) | Sonnet 5 (Sep 1) | GPT-5.6 Terra |
|---|---|---|---|---|
| 5M | $37.50 | $50.00 | $75.00 | $75.00 |
| 10M | $75.00 | $100.00 | $150.00 | $150.00 |
| 50M | $375.00 | $500.00 | $750.00 | $750.00 |
Rates per Google, TechCrunch, and OpenRouter; input tokens billed on top.
At 50 million output tokens a month, the gap is $375 versus $750, a real line item. That is the sticker read. The token-efficiency claim, if it holds for your workload, widens it.
Google says Gemini 3.6 Flash cuts output tokens by up to 65% on DeepSWE, a long-horizon software-engineering benchmark, where its score also climbed from 37% to 49%, per VentureBeat and MarkTechPost. Read that number carefully. The 65% is the best case on one agentic coding benchmark, not a blanket discount on every call. The broad figure is smaller. Artificial Analysis, measuring across general tasks, put the output-token reduction near 17%, per MarkTechPost. Budget on 17%. Treat 65% as the ceiling that step-heavy coding agents occasionally hit.
What does that do to the 5-million-token workflow? A 17% token reduction on comparable work pulls the effective spend from $37.50 toward roughly $31, because you generate fewer tokens to finish the same task. Take a single coding-agent step that burns 1,000 output tokens on the older Gemini 3.5 Flash. That step cost $0.009 at 3.5 Flash's $9.00 output rate, per MarkTechPost. If 3.6 Flash finishes it in about 350 tokens at the same quality, the step drops to roughly $0.0026 at the new $7.50 rate. Cheaper model, fewer tokens, both pulling the same direction. The candid caveat: only tasks shaped like DeepSWE see the 65% end of that. Summarization and classification see closer to the 17%.
Where does the math stop favoring Gemini? On tasks that need one careful answer rather than a fast one. If Sonnet 5 nails a legal summary in a single pass and Gemini 3.6 Flash needs a retry or a longer prompt to match it, the token count on Gemini climbs and the quality edge on Sonnet earns its price. The efficiency claim is about coding-agent step counts, not about being right on hard reasoning the first time.
What makes Gemini 3.6 Flash different from 3.5 Flash
Two things moved at once. Google dropped the output rate from $9.00 on Gemini 3.5 Flash to $7.50 on 3.6 Flash while making the model generate fewer tokens per task, per MarkTechPost. Lower verbosity plus a lower price compounds.
The knob that matters most for operators is configurable reasoning effort. Gemini 3.6 Flash lets you set how hard the model thinks per request, and it supports parallel tool use across complex workflows, per GitHub's changelog. In the raw API, that control is a thinking budget you set in generationConfig. Turn it to zero for cheap, fast calls like routing and tagging:
# Fast path: reasoning off, for classification and routing
curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.6-flash:generateContent?key=$GEMINI_API_KEY" \
-H 'Content-Type: application/json' \
-d '{
"contents": [{"parts": [{"text": "Classify this lead as sales, support, or billing: <email body>"}]}],
"generationConfig": {"thinkingConfig": {"thinkingBudget": 0}}
}'Raise the budget for the occasional hard task, same model, same API client, no second connection to maintain:
# Harder path: give the model room to reason on a multi-step task
curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.6-flash:generateContent?key=$GEMINI_API_KEY" \
-H 'Content-Type: application/json' \
-d '{
"contents": [{"parts": [{"text": "Draft a three-step follow-up plan for this stalled deal: <deal notes>"}]}],
"generationConfig": {"thinkingConfig": {"thinkingBudget": 4096}}
}'That one-model-two-modes property is a genuine differentiator for tiered workloads. Sonnet 5 has no public reasoning dial, so covering both cheap and hard calls with Claude means swapping to a heavier model. Terra exposes a reasoning parameter, per Finout. If your automation mixes ten thousand cheap classifications with a few hundred harder drafts a day, one Gemini connection with a per-request budget cuts the operational overhead of maintaining two model integrations.
Two facts to plan around. Gemini 3.6 Flash and 3.5 Flash-Lite carry a knowledge cutoff of March 2026, per OfficeChai, so anything after that date has to come through tool calls or retrieval, not the model's memory. And 3.6 Flash is not the Pro tier Google promised for June. Google's Gemini 3.5 Pro underperformed on coding benchmarks and went back into retraining in late June without published results, per TechCrunch. If you were waiting on a Pro-class Gemini to challenge the top of the Claude and OpenAI lineups, this launch is not it. It is a better, cheaper Flash.
Gemini 3.6 Flash in GitHub Copilot
You do not need to touch the API to use it. The model is rolling out in GitHub Copilot's model picker for Pro, Pro+, Max, Business, and Enterprise, per GitHub's changelog. It shows up in Visual Studio Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the Copilot app, and the Copilot cloud agent. In early testing, Google reported higher task-completion rates and better token efficiency than Gemini 3.5 Flash across coding and agentic work, per the same changelog.
Two setup notes. On Business and Enterprise plans, an admin has to enable the Gemini 3.6 Flash Preview policy before anyone can select it. And the model bills at provider list pricing under Copilot's usage-based billing, per GitHub's changelog, so the same $7.50 output rate applies to your seat's request budget. Copilot access itself carries no Pondero commission; the GitHub Copilot link is here for completeness, not because we earn on it.
When do you pick Gemini 3.6 Flash over Sonnet 5 inside Copilot? For fast iterative completions and cost-sensitive team billing, the token efficiency shows up directly on the usage meter. For a hard refactor across a large repo where one correct pass beats three fast tries, Sonnet 5 still earns the request. If you run Cursor instead of Copilot, the July 2026 Cursor review covers which models are wired in there.
Which model for which workflow
This is the table to screenshot. Costs assume roughly 5 million output tokens a month at each model's list output rate above.
| Use case | Pick | Why | Rough monthly output cost |
|---|---|---|---|
| High-volume classification (email routing, tagging, moderation) | Gemini 3.6 Flash | Cheapest output, fast, reasoning budget set to zero | ~$37.50 |
| Content draft generation (newsletters, reports, social) | Claude Sonnet 5 | Quality edge at intro pricing; budget the Sep 1 step to $15/MTok | ~$50, then ~$75 |
| Multi-step reasoning (legal analysis, complex code review) | Fable 5 or GPT-5.6 Sol | Not this tier; see the Fable 5 guide | Higher tier |
| Long-context documents (past ~200K tokens) | GPT-5.6 Terra | 1M-token context window, per OpenRouter | ~$75 |
| Make automation (an LLM call per trigger) | Gemini 3.6 Flash for volume, Sonnet 5 for quality | Cheapest per-call entry; quality when the draft ships to a customer | ~$18.75 to ~$50 |
| GitHub Copilot daily coding | Gemini 3.6 Flash or Sonnet 5 | Flash for cost-sensitive teams, Sonnet for hard refactors | Provider list rate |
Terra's edge is the context window. At 1 million tokens, it swallows a document set that Flash and Sonnet 5 would need chunking to handle, per OpenRouter. Stuffing entire codebases or long contract bundles into a single call makes that $15 output rate worth it. Below that, you are paying Terra's premium for headroom you never use.
The Make angle: which model on which scenario
Make (formerly Integromat) calls the Google AI Studio API, the Claude API, and the OpenAI API from inside one scenario, so the model choice is a per-module decision, not a platform lock-in. That makes it a clean place to route the volume work to Gemini 3.6 Flash and reserve Sonnet 5 for the modules whose output a human sees.
Here is a concrete scenario. A Make automation fires on every new form submission, classifies the lead, and drafts a follow-up email. At 5,000 submissions a month and 500 output tokens each, that is 2.5 million output tokens. On Gemini 3.6 Flash, $18.75. On Sonnet 5 from September 1, $37.50. If the classification step and the draft are both Flash, you pay the lower number; if you promote just the customer-facing draft to Sonnet 5, you split the difference and only pay the premium where it lands in someone's inbox. In Make, that is an HTTP module posting to the Gemini endpoint:
{
"url": "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.6-flash:generateContent?key={{GEMINI_API_KEY}}",
"method": "POST",
"headers": [{ "name": "Content-Type", "value": "application/json" }],
"body": {
"contents": [{ "parts": [{ "text": "Route this lead and draft a reply: {{1.form_body}}" }] }],
"generationConfig": { "thinkingConfig": { "thinkingBudget": 0 } }
}
}Make's AI modules support all three providers in a single scenario, so you can start with Make and swap models per step without rebuilding the flow. If a scenario scrapes competitor pricing or product pages before the LLM call, Firecrawl feeds clean page content straight into Gemini 3.6 Flash, and Flash's lower token cost is what makes a scraper-plus-model loop cheap enough to run on every page. For a fuller build, the Make AI agents guide walks the setup end to end.
The Claude Sonnet 5 window closes August 31
Running Sonnet 5 at intro pricing and watching the bill climb? Benchmark your token usage this week. From September 1, Sonnet 5 output costs 50% more, $10 to $15 per MTok, and input rises the same 50%, per TechCrunch. On the 5-million-token workflow, that is $50 now against $75 in September.
The Flash launch changes the September read. At Sonnet 5's standard rate, Flash output is 50% cheaper per token, $7.50 against $15.00, per Google's pricing and TechCrunch. Whether that swap is worth it comes down to task type, and this is where the honest split lives: for classification, routing, and summarization, Flash almost certainly holds quality and you pocket the difference; for reasoning-heavy drafting, Sonnet 5 likely still earns its rate. Run the comparison on your own traffic before the intro window shuts, not after the invoice jumps.
The verdict
Route your workloads like this.
For high-volume classification, routing, and Make scenarios where a small quality gap is fine, use Gemini 3.6 Flash. It has the cheapest output at $7.50/MTok, per Google's pricing, the token-efficiency tailwind on agentic work, and a reasoning budget you dial to zero for the cheap calls. Wire it into Make and it covers most of an automation's LLM steps at the lowest cost of the three.
For draft generation and reasoning tasks where quality is worth paying for, use Claude Sonnet 5 while the intro rate holds through August 31. The candid con is the September 1 step to $15 output, per TechCrunch, which is exactly when Gemini 3.6 Flash starts looking like the default for anything that is not quality-critical.
For workloads that need a million-token context window, use GPT-5.6 Terra. You are paying $15 output for the context headroom, per OpenRouter, so only pick it when you actually fill that window; otherwise Luna at $6 (Finout) or Gemini 3.6 Flash is the smarter spend.
For the hardest multi-step reasoning and agentic coding, none of these three is the right tool. That is Fable 5 and GPT-5.6 Sol territory, and the Fable 5 vs Sonnet 5 pricing guide has the tier-by-tier breakdown.
FAQ
Does Gemini 3.6 Flash support function calling? Yes. It supports parallel tool use across complex workflows, per GitHub's changelog, which is the capability agent frameworks lean on to call multiple tools in one turn.
Can I use Gemini 3.6 Flash outside Google AI Studio? It is available through the Gemini API and rolling out in GitHub Copilot's model picker across VS Code, JetBrains, Xcode, the Copilot CLI, and the cloud agent, per GitHub's changelog. Copilot access needs an admin to enable the Gemini 3.6 Flash Preview policy on Business and Enterprise plans first.
Is Claude Sonnet 5 available in Make? Make calls the Claude API directly, so you can point an HTTP or AI module at Sonnet 5 the same way you would at Gemini. The model choice is per-module, which is why routing cheap steps to Flash and quality steps to Sonnet 5 in one scenario works.
What happens to my Gemini 3.5 Flash workflows if I switch to 3.6? The output price drops from $9.00 to $7.50/MTok and the model tends to generate fewer tokens per task, per MarkTechPost. Swap the model string, re-check any prompt that depended on the older model's verbosity, and re-run your evals before you flip production.
Will GPT-5.6 Terra get cheaper as it matures? No one has announced a cut. Terra released July 9, 2026 at $2.50/$15, per OpenRouter. If you need OpenAI but not the full context window, Luna at $1/$6 is the cheaper tier today, per Finout.