Short answer. At the balanced tier, Claude Sonnet 5.5 and GPT-6.2 Sol both list $2 input / $10 output per million tokens; Gemini 4 Argon lists $2 / $10 through 30 November 2026, then $4 / $20. At the low-cost tier, Claude Haiku 5.5 and GPT-6 Luna list $0.10 / $0.50; Gemini 3.5 Flash-Lite lists $0.30 / $2.50. Claude 4.6+ and Gemini 4 list a 1M-token context; OpenAI prices short context up to 272K separately from long context.
Current API tiers
| Tier | Claude | Gemini | GPT |
|---|---|---|---|
| Flagship | Fable 5.2 — $8 / $40 | Gemini 4 Carbon — $2 / $10 (to 30 Nov; then $4 / $20) | GPT-6.2 Astra — $8 / $40 |
| Reasoning / balanced | Opus 5.5 — $4 / $20 Sonnet 5.5 — $2 / $10 | Gemini 4 Argon — $2 / $10 (to 30 Nov; then $4 / $20) Neon — $3 / $15 | GPT-6.2 Sol — $2 / $10 GPT-6.1 Terra — $1 / $5 |
| Low cost | Haiku 5.5 — $0.10 / $0.50 up to 100K; $0.50 / $2.50 above | Gemini 3.8 Flash — $0.75 / $3.75 Flash-Lite — $0.30 / $2.50 | GPT-6 Luna — $0.10 / $0.50 |
| Context | 1M on 4.6+ | 1M on Gemini 4 | Short ≤272K; long >272K at a higher rate |
Prices per 1M input/output tokens, checked 20 November 2026 against Anthropic, Google Gemini and OpenAI pricing documentation. Promotional rates expire; confirm before budgeting.
Cost at 10M tokens
Assume a 70/30 input/output mix:
Claude Sonnet 5.5: 7M × $2 + 3M × $10 = $14 + $30 = $44
Claude Haiku 5.5: 7M × $0.10 + 3M × $0.50 = $0.70 + $1.50 = $2.20
Gemini 4 Argon: 7M × $2 + 3M × $10 = $44 (to 30 Nov; then $88)
Gemini 3.8 Flash: 7M × $0.75 + 3M × $3.75 = $5.25 + $11.25 = $16.50
Flash-Lite: 7M × $0.30 + 3M × $2.50 = $2.10 + $7.50 = $9.60
GPT-6.2 Sol: 7M × $2 + 3M × $10 = $44
GPT-6 Luna: 7M × $0.10 + 3M × $0.50 = $2.20
This is the list-price view. It ignores caching, batch discounts, long-context surcharges and retries. A model with a 10% higher token price but a 20% lower retry rate can be cheaper per completed task.
Where each family is strongest
| Workload | What to test first | Why |
|---|---|---|
| Long-form writing and editing | Claude Sonnet/Opus, GPT Sol | Instruction-following and tone consistency; run a blind review |
| Repository-level coding | Claude Sonnet/Opus, GPT Sol, Gemini Argon | Tool use, diff quality and test pass rate on your own repo |
| Long documents | Claude 4.6+, Gemini 4, GPT long-context | 1M context vs OpenAI's long-context pricing tier |
| High-volume classification | Haiku 5.5, GPT-6 Luna, Flash-Lite | Low per-token cost; accuracy on your label set |
| Multimodal | Gemini 4, GPT multimodal, Claude vision | Image, PDF and video support differ by model |
| Agents with tools | All three; test tool-call correctness | Schema adherence, parallel calls and error recovery |
Tool use and structured output
All three support tool calling and JSON output, but the details differ: schema strictness, parallel tool calls, streaming events and error shapes. Build a 30-case tool suite and measure:
- Correct tool selected and arguments valid.
- JSON parses without repair and matches the schema.
- Model recovers from a tool error instead of looping.
- Latency and token overhead of the tool definitions.
Data terms
Paid API tiers generally do not train on customer data by default, but the terms are not identical. Anthropic, OpenAI and Google each publish data-use and retention terms, and free consumer tiers differ from paid API tiers. For regulated workloads, confirm retention, residency, zero-data-retention options and subprocessors in the contract — not in a blog post.
Migration path
- Abstract the client: one interface for chat, tools and JSON; provider adapters underneath.
- Map model tiers, not names: flagship, balanced, cheap. Pin exact model IDs per environment.
- Re-test tools and JSON: schema adherence is the most common migration failure.
- Run a parallel eval: 20–50 real tasks per provider, scored on correctness, latency and cost per solved task.
- Keep a fallback: a second provider or a multi-model API for rate-limit and outage resilience.
When a multi-model API wins
If you do not want three contracts, three billing systems and three SDKs, use an OpenAI-compatible API that routes across model families. Plugsky's Hobby is $8/month, Starter $20, Builder $60 and Scale $120; Builder and Scale have no monthly token cap, with fair-use concurrency and RPM. See pricing and the model catalogue.
FAQ
Which is cheaper: Claude, Gemini or GPT?
At the balanced tier, Claude Sonnet 5.5 and GPT-6.2 Sol both list $2/$10; Gemini 4 Argon lists $2/$10 through 30 November, then $4/$20. At the low tier, Haiku 5.5 and GPT-6 Luna list $0.10/$0.50.
Which has the largest context window?
Claude 4.6+ and Gemini 4 list 1M tokens. OpenAI prices short context up to 272K separately from long context above 272K.
Which is best for coding?
No universal winner. Run a repository-level eval with 20–50 real tasks and measure pass rate, diff quality and cost per solved task.
Do these APIs train on my data?
Paid API tiers generally do not train on customer data by default, but terms differ by vendor, tier and region. Review the current data-use terms.
Can I switch without rewriting?
Similar request shapes, different tool/JSON/streaming details. An OpenAI-compatible abstraction reduces the change to a base URL and model ID, but re-test tools and structured output.
Reviewed 20 November 2026. Prices and context windows were checked against Anthropic, Google Gemini and OpenAI API pricing documentation on that date. Promotional rates expire; confirm before committing. This page is editorial and does not rank providers by payment.