Canonical: https://botbento.com/blog/model-cost-comparison-bot-tool-calls-nov-2026/
Format: Markdown representation of the public HTML page.

[Home](/) / [Blog](/blog/)

FIELD NOTES / 6 MIN READ

# Claude, GPT-6, or Gemini 3.1 Pro: Which Model Is Cheapest for Your Bot's Tool Calls?

Three providers now sell a mid-tier model at close to the same headline price, so picking the cheapest model for a tool-calling bot comes down to output ratio, cache hits, and whether your prompts cross a long-context threshold. This piece runs one illustrative task through Anthropic's, OpenAI's, and Google's current official rate cards to show where the numbers actually diverge.

By BotBento Editorial · Published 2026-11-11 · Updated 2026-11-11

AI-assisted editorial: researched and drafted with AI using the primary sources linked below. Check the cited evidence and any stated test scope. Illustrative examples are not measured product results. [Editorial policy](/editorial/).

## In this article

- [The short answer](#the-direct-answer)
- [What each provider actually charges right now](#current-rate-cards)
- [The API you're forced into changes the comparison](#api-shape-matters-too)
- [A worked example: one tool call per run, 1,000 runs a month](#worked-illustrative-example)
- [What actually moves the bill in production](#what-changes-the-real-bill)
- [The decision rule](#the-decision)[Read as Markdown ](/text/blog/model-cost-comparison-bot-tool-calls-nov-2026/index.md)

## Key takeaways

- Claude Sonnet 5.5 and GPT-6.1 Sol both list $2 input / $10 output per million tokens, so for most small-team tool-calling workloads the tie-breaker is cache-hit rate and API shape, not sticker price.
- Gemini 3.1 Pro's $12-per-million output rate (versus $10 on Sonnet 5.5 and Sol) makes it slightly more expensive for output-heavy tool results, and it doubles to $4/$18 once a prompt passes 200,000 tokens.
- Every tool schema you attach is billed as input on every call whether the model uses it or not, so trimming unused tools cuts cost before any model swap does.

## The short answer

For a bot that calls one or two tools per run and sends a few thousand tokens of context each time, Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol currently cost the same per token: $2 per million input tokens and $10 per million output tokens. Gemini 3.1 Pro is close but not identical — it charges $2 input and $12 output per million tokens for prompts up to 200,000 tokens. On a typical short tool-calling turn, that gap is small in absolute dollars but it compounds on output-heavy workloads, like a bot that drafts long replies after a tool call returns data.

None of the three is a clear winner on price alone. The real cost driver for most bot builders is not the per-token rate but three mechanics that behave differently across providers: how much your attached tool definitions add to every request, how cache hits are priced, and what happens once a prompt crosses a long-context threshold.

## What each provider actually charges right now

Anthropic's current Claude lineup lists Sonnet 5.5 and Sonnet 5 at the same standard rate, with cache reads priced as a 5 percent multiplier on the newest mid and top tiers.

- Claude Sonnet 5.5: $2 input / $10 output per million tokens, cache reads at 5% of input ($0.10/MTok)
- Claude Opus 5.5: $4 input / $20 output per million tokens, cache reads at 5% of input ($0.20/MTok)
- OpenAI GPT-6.1 Sol: $2 input / $10 output per million tokens short-context, cached input $0.10/MTok, rising to $4/$15 above 272K input tokens
- OpenAI GPT-6.1 Terra: $1 input / $5 output per million tokens, cached input $0.05/MTok
- OpenAI GPT-6 Astra: $10 input / $50 output per million tokens, cached input $1/MTok
- OpenAI GPT-6 Luna: $0.10 input / $0.50 output per million tokens, cached input $0.01/MTok
- Gemini 3.1 Pro: $2 input / $12 output per million tokens up to 200K-token prompts, stepping to $4/$18 above that threshold

Sources: [Pricing - Claude Platform Docs](https://platform.claude.com/docs/en/about-claude/pricing), [Pricing | OpenAI API](https://developers.openai.com/api/docs/pricing), [Gemini Developer API pricing | Gemini API | Google AI for Developers](https://ai.google.dev/gemini-api/docs/pricing).

## The API you're forced into changes the comparison

Price parity between Claude Sonnet 5.5 and GPT-6.1 Sol doesn't mean the integration cost is identical. OpenAI's current flagship tiers only support tool calling through one interface: GPT-6 Astra and GPT-6.1 Sol require the Responses API for tool calling, so a bot still wired to Chat Completions for function calls has to migrate before it can use either model. That's a one-time engineering cost, not a per-token one, but it belongs in the decision if you're already running a stable integration on an older model.

Gemini's tool billing has its own quirk worth knowing before you budget: every token the model processes is counted and billed, including thinking tokens and internal tool system instructions and function declarations when using grounding. In practice this means a bot's tool schema and any grounding instructions are charged as input on every call, the same way OpenAI and Anthropic charge for attached tool definitions — the difference is that Google's documentation calls this out explicitly for grounding-enabled requests, so check your own tool count against the usage field rather than assuming the visible prompt is the whole bill.

Sources: [Function calling | OpenAI API](https://developers.openai.com/api/docs/guides/function-calling), [Gemini Developer API pricing | Gemini API | Google AI for Developers](https://ai.google.dev/gemini-api/docs/pricing).

## A worked example: one tool call per run, 1,000 runs a month

This is an illustrative calculation, not a BotBento-run benchmark. Assume a small-team bot that looks up a calendar slot through one tool once per run: roughly 2,000 input tokens (system prompt, tool schema, and user message) and 300 output tokens per run, executed 1,000 times a month with no cache hits. Multiplying each provider's listed rate by that volume gives a rough monthly floor, excluding retries, thinking tokens, and any multi-step tool loop.

At those volumes, Claude Sonnet 5.5 and GPT-6.1 Sol both land near $7 a month (2 million input tokens at $2/MTok plus 300,000 output tokens at $10/MTok). Gemini 3.1 Pro comes out slightly higher, close to $7.60, purely because of its $12 output rate. GPT-6.1 Terra undercuts all three at roughly $3.50 because its $1/$5 rate is half of Sol's. GPT-6 Astra, the OpenAI flagship, runs far higher — around $35 for the same volume — because its $10/$50 rate is five times Sol's on both sides. None of these totals include reasoning or thinking tokens, which on every current reasoning model bill at the output rate and can multiply the output-token count several times over for a single turn.

The lesson from the math isn't that one model is universally cheapest. It's that the gap between a mid-tier model (Sonnet 5.5, Sol, Terra, Gemini 3.1 Pro) and a flagship (Opus 5.5, Astra) is far larger than the gap between mid-tier models from different vendors. Picking the right tier matters more than picking the right vendor for most tool-calling bots.

## What actually moves the bill in production

Caching is the biggest lever if your tool schema and system prompt stay stable across calls. Anthropic prices a cache hit on Opus 5.5 and Sonnet 5.5 at 5 percent of the standard input price, which is cheaper than the 10 percent most other Claude models charge for a cache read. A cache write, by contrast, costs more than an uncached call the first time — 1.25x input for a 5-minute cache and 2x for a 1-hour cache — so caching only pays off once you're reading the cached block at least once (5-minute) or twice (1-hour).

Long-context thresholds are the other trap. GPT-6.1 Sol's rate doubles on input and rises 50 percent on output once a prompt passes 272,000 input tokens. Claude Haiku 5.5 has its own step-up above 100,000 tokens, while the Sonnet and Opus 5.5 tiers bill the full 1-million-token window at the standard rate with no surcharge — a structural difference worth checking before you assume every Claude tier behaves the same way at long context. Gemini 3.1 Pro re-rates the entire request, input and output both, once combined prompt and output crosses 200,000 tokens, which is a lower threshold than either competitor and a cliff rather than a gradual increase.

For a tool-calling bot specifically, the most overlooked cost is the tool schema itself. Every function or tool definition attached to a request is billed as input tokens on every single call, regardless of whether the model ends up using that tool. A bot carrying ten unused tool definitions pays for ten tool definitions' worth of tokens on every turn. Anthropic's own documentation shows this concretely for its built-in computer-use toolset: declaring the toolset with its default members adds about 4,500 input tokens to a request before the model does anything, and declaring the browser-use toolset adds about 6,600. Custom function tools follow the same pattern at a smaller scale — trimming an unused tool from the list is often a bigger saving than switching models.

Sources: [Pricing - Claude Platform Docs](https://platform.claude.com/docs/en/about-claude/pricing).

## The decision rule

If your bot's tool-calling workload is input-heavy (long context, short replies) and you're already integrated with one provider's API shape, the rate parity between Claude Sonnet 5.5 and GPT-6.1 Sol means a model swap on price grounds alone is rarely worth the migration cost. If your workload is output-heavy — long drafted replies, multi-step tool chains that each generate text — the $10 versus $12 output gap between Sol/Sonnet and Gemini 3.1 Pro adds up faster than the input-side numbers suggest, and GPT-6.1 Terra's $5 output rate is worth testing first if your task tolerates a smaller model.

Before comparing vendors, cut your own tool list to only what a given task needs, measure your actual cache-hit rate against the write cost, and check whether your typical prompt length sits comfortably under each provider's long-context threshold. Those three checks change a bot's real monthly bill more than any single model choice does, and they're measurable from your own usage logs rather than a vendor's rate card.

## Primary sources

Sources checked 2026-11-11. Standards and product documentation can change; follow the linked version when implementing.

- [Pricing - Claude Platform Docs](https://platform.claude.com/docs/en/about-claude/pricing) — Anthropic
- [Pricing | OpenAI API](https://developers.openai.com/api/docs/pricing) — OpenAI
- [Function calling | OpenAI API](https://developers.openai.com/api/docs/guides/function-calling) — OpenAI
- [Gemini Developer API pricing | Gemini API | Google AI for Developers](https://ai.google.dev/gemini-api/docs/pricing) — Google

BotBento is in development. [Suggest a correction](/contact/).
