Gemini Code Assist and CLI thinking-token costs: how to stop a coding session from burning $100 a day (2026)

Summary. Gemini bills reasoning as output. Every model on Google's pricing page lists its output rate as "output price (including thinking tokens)," so the background thinking a model does before it answers is charged at the full output rate. On Gemini 3.1 Pro Preview that output rate is $12.00 per million tokens for prompts up to 200k, rising to $18.00 above 200k, while input is $2.00. A reasoning-heavy coding session emits far more output than the code you see, which is how developers on Google's AI developer forum have reported daily costs of $100 to $140 after switching to thinking models. Two things changed the stakes on June 18, 2026: the free and individual Gemini Code Assist tiers, plus Google AI Pro and Ultra access inside the IDE extensions and Gemini CLI, stopped working, leaving paid Standard at $19 per user per month and Enterprise at $45. So the spend is now yours to manage. This guide explains how the billing works and how to cap it with model choice, the thinking-control config, context caching and budgets.

Why the bill moved

A thinking token is a token the model generates while reasoning through a problem before it writes its visible answer. You do not see these tokens in the response, but you pay for them, because Gemini folds them into the output-token count. On a coding assistant that reasons about your repository on every completion and every terminal command, those hidden tokens accumulate on each call.

The rate gap between model tiers is the whole game. The table shows the Standard paid rates for prompts up to 200k tokens.

Read the output column, not the input column, because thinking lands there. Gemini 3.1 Pro output at $12.00 per million is about 8 times the $1.50 that Gemini 3.1 Flash-Lite charges, and prompts above 200k tokens push Pro output to $18.00. Point a coding agent at Pro with reasoning on, let it churn through a large repository context, and the output line is where the money goes.

The arithmetic of a burned afternoon

Put numbers on it. Say one heavy session, refactoring across several files with the model reasoning on each step, emits 3 million output tokens counting thinking. On Gemini 3.1 Pro at $12.00 per million that session costs about $36 in output alone. The same 3 million tokens on Gemini 3.5 Flash-Lite at $2.50 per million costs about $7.50, and on Gemini 3.1 Flash-Lite at $1.50 it costs about $4.50. Run three or four Pro sessions like that in a day and you are at the $100-plus figure the forums describe, without doing anything that felt unusual.

These are illustrative figures, not a quoted benchmark; your real numbers depend on how much the model reasons and how large your context is. The point holds regardless of the exact token count: the model tier multiplies every session, so tier choice is the single biggest lever you have.

Control lever 1: pick the cheaper model for routine work

Most coding tasks, code completion, boilerplate, renaming, small edits, do not need frontier reasoning. Route those to Flash-Lite and keep Pro for genuine architectural work. Because output including thinking is where the cost sits, moving routine calls from $12.00 to $1.50 per million output is a real cut, not a rounding change. Our budget LLM tier cost comparison and the Gemini 3.6 Flash token-efficiency guide walk through where each tier earns its price.

Control lever 2: the thinking config

Gemini exposes reasoning controls through thinkingConfig, but the exact knob depends on the model generation, and this trips people up.

The sharp edges are worth stating plainly. On Gemini 2.5 Flash you can set thinkingBudget: 0 and turn reasoning off entirely. On Gemini 2.5 Pro you cannot fully disable it; the budget floor is 128 tokens. On Gemini 3 and 3.1 you use thinkingLevel rather than a raw budget, and you cannot switch thinking fully off on 3.1 Pro. There is also a failure mode: because thinking tokens count against maxOutputTokens, a limit set too low can be consumed entirely by thinking and return an empty response, a bug developers have logged against the 2.5 and 3 Flash models. Size maxOutputTokens for thinking plus answer, not answer alone.

Control lever 3: cache context and watch grounding

A coding assistant resends large chunks of the same repository context on every call. Context caching prices that repeated input far lower than fresh input, so caching the stable parts of a prompt cuts the input line on high-volume sessions. Separately, if your setup uses grounding with Google Search, Gemini 3 gives 5,000 grounded prompts per month free and then charges $14 per 1,000 search queries, and one request can fire more than one query. Track it, because it is billed on top of tokens. For teams running several models, our LLM hybrid-routing spend framework covers deciding which calls go where.

Control lever 4: budgets, quotas and seat math

After June 18, 2026 the free ride ended. The tiers now look like this.

The seat fee is the predictable part; the usage-based API calls billed at the underlying model's token rate on top of the seat fee are the part that surprises finance. Set per-project budget alerts in Google Cloud, cap keys with quotas, and default your team's tooling to a Flash-tier model so the expensive Pro path is a deliberate choice, not the resting state. The same discipline applies to the Claude Sonnet 5 tokenizer cost cliff, where hidden token growth also moves the bill.

Control lever 5: scope the context you send

Prompt size is a rate multiplier, not just a token count. On Gemini 3.1 Pro, crossing 200k tokens in a prompt lifts output from $12.00 to $18.00 per million and input from $2.00 to $4.00, so a coding agent that stuffs an entire repository into every request pays the higher tier on work a scoped context would have kept cheaper. Larger context also gives the model more to reason over, which inflates thinking tokens on top of the input charge.

The fix is discipline about what the assistant sees. Feed the files and symbols relevant to the task rather than the whole tree, lean on the tool's file-scoping features, and split a sprawling refactor into smaller scoped requests instead of one giant-context call. Smaller prompts cut the input line, keep you under the 200k tier break, and reduce the reasoning surface that drives the output line. It is the least glamorous lever and often the most effective, because it attacks input and thinking cost at the same time.

India-specific considerations

For Indian teams the seat and usage both bill in dollars, so the weak rupee lifts the real cost. At roughly 96 rupees to the dollar in late July 2026, Standard at $19 per user is about 1,824 rupees per user per month, and Enterprise at $45 is about 4,320 rupees, before any usage. On usage, a single $36 Pro session works out near 3,456 rupees, versus about 720 rupees for the same session on Flash-Lite. For a ten-developer team, defaulting routine work to a Flash tier and reserving Pro for hard problems is the difference between a predictable and an alarming monthly invoice. Budget in rupees and set the Google Cloud alerts before you scale seats.

FAQ

How eCorpIT can help

eCorpIT helps engineering teams get Gemini Code Assist and CLI spend under control without slowing developers down. We set model-routing defaults so routine work runs on a Flash tier, apply the right thinkingConfig per model, cache repository context, and wire Google Cloud budget alerts and quotas so a runaway session is caught early. That work sits inside our LLM migration and cost optimization service. To review your AI coding spend, contact eCorpIT.

References

1. Google AI for Developers, "Gemini Developer API pricing": ai.google.dev/gemini-api/docs/pricing

1. Google AI for Developers, "Gemini thinking" (thinkingConfig, thinkingBudget, thinkingLevel): ai.google.dev/gemini-api/docs/thinking

1. Google for Developers, "Gemini Code Assist consumer accounts" (deprecation): developers.google.com/gemini-code-assist

1. getDX, "AI coding assistant pricing and ROI guide (2026)": getdx.com/blog/ai-coding-assistant-pricing

1. CloudZero, "Gemini pricing in 2026: every model, every plan, and the thinking tokens nobody budgeted for": cloudzero.com/blog/gemini-pricing

1. Finout, "Gemini Pricing in 2026 for Individuals, Orgs & Developers": finout.io/blog/gemini-pricing-in-2026

1. cline (GitHub), "thinkingBudget defaults to 0, incompatible with Gemini 2.5 Pro": github.com/cline/cline/issues/7735

1. ha-llmvision (GitHub), "Gemini 2.5/3 Flash thinking tokens consume maxOutputTokens, causing empty responses": github.com/valentinfrlch/ha-llmvision/issues/609

1. ofox.ai, "Gemini CLI Free Tier Shut Down: fixes that work (2026)": ofox.ai/blog/gemini-cli-free-tier-shutdown-fix-2026

1. Exchange Rates UK, "US Dollar to Indian Rupee spot exchange rates history 2026": exchangerates.org.uk

_Last updated: July 29, 2026._