TL;DR (May 2026): Gemini 3.1 Pro is the cheapest flagship at $2/$12 per 1M tokens, 60% under Claude Opus 4.7 ($5/$25). GPT-5.5 is the new top tier at $5/$30 with the best tool-use scores. Prompt caching cuts up to 90% off all three. The decision is rarely "cheapest", it's "which capability per dollar fits your workload."
Three months ago this comparison would have been cleaner. GPT-5, Claude Opus 4.6, Gemini 3 Pro. Three flagships, similar order of magnitude, you pick the one your team likes. May 2026 is messier.
OpenAI shipped GPT-5.5 in late April. Anthropic moved Claude Opus 4.6 to 4.7 with a price hold. Google's Gemini 3.1 Pro replaced 3 Pro and the free tier dynamics shifted on April 1. The pricing math that worked in February is wrong now. Here's the May 2026 version.
The price card, current as of May 23, 2026
| Model | Input ($/1M) | Output ($/1M) | Cached input ($/1M) | Context window |
|---|---|---|---|---|
| GPT-5.5 | $5.00 | $30.00 | $0.50 | 270K |
| GPT-5.2 | $1.75 | $14.00 | $0.18 | 200K |
| Claude Opus 4.7 | $5.00 | $25.00 | $0.50 (with prompt caching) | 200K |
| Gemini 3.1 Pro (≤200K ctx) | $2.00 | $12.00 | $0.20 | 2M |
| Gemini 3.1 Pro (>200K ctx) | $4.00 | $18.00 | $0.40 | 2M |
Three things that jump out and shouldn't be ignored:
- Gemini wins flagship pricing by a lot. 60% cheaper than Claude Opus 4.7 on input, 52% cheaper on output. The price gap was already wide in Q1; the 3.1 release widened it.
- The cache discount is enormous. 90% off across all three providers for repeated context. Most production workloads have a stable system prompt and a stable RAG corpus, if you're not using prompt caching, you're paying full price for the same tokens over and over.
- Context window is no longer a tiebreaker. Gemini's 2M window is 7-10x what GPT and Claude offer, but very few workloads actually need it. If yours doesn't, the context window stops mattering for the choice.
Real-world cost: three workloads compared
Headline per-token pricing is the wrong lens because no real workload hits exactly the headline shape. Three realistic patterns, with the math worked out:
Workload A: support agent (small input, small output, high volume)
Average request: 2K input tokens, 500 output tokens, 50K requests/month. The input includes a 1K system prompt that's identical across requests, so prompt caching applies to half the input.
| Model | Monthly cost (no caching) | Monthly cost (with caching) |
|---|---|---|
| Claude Opus 4.7 | $1,125 | $687 |
| GPT-5.5 | $1,250 | $762 |
| GPT-5.2 | $525 | $397 |
| Gemini 3.1 Pro | $500 | $350 |
On a workload like this, Gemini 3.1 Pro is roughly half the cost of Claude Opus and a third the cost of GPT-5.5. Use Opus or GPT-5.5 here only if the quality delta on your specific task is large enough to justify 2x-3x the bill. Most support agents don't see that delta.
Workload B: long-context document analysis (huge input, small output, low volume)
Average request: 150K input tokens, 2K output tokens, 1,000 requests/month. No repeated system prompt, each document is unique. Caching doesn't help here.
| Model | Monthly cost |
|---|---|
| Claude Opus 4.7 | $800 |
| GPT-5.5 | $810 |
| Gemini 3.1 Pro | $324 |
Gemini's pricing edge gets larger on long-context workloads, and the 2M context window means you can pack more into a single call instead of chunking. Opus 4.7 and GPT-5.5 are tied at this end. Pick on quality, not cost.
Workload C: agent loop with tool calls (medium tokens, many turns per task)
Average task: 10 turns, each turn 5K input / 1K output, 5,000 tasks/month. The input on later turns includes the conversation history, so caching helps massively.
| Model | Cost without caching | Cost with prompt caching |
|---|---|---|
| Claude Opus 4.7 | $3,500 | $1,050 |
| GPT-5.5 | $3,750 | $1,125 |
| Gemini 3.1 Pro | $1,400 | $420 |
The caching discount on agent loops is enormous because each turn re-sends the previous turns. If you're running a Claude or GPT agent loop and not paying attention to caching, your bill is probably 3x what it could be.
Pick by capability, not just price
Cost is real but it's not the whole picture. Where each flagship actually wins on capability, based on the workloads we've watched teams run in 2026:
Claude Opus 4.7 wins at
- Code generation longer than ~500 lines. Cohesion and structure stay tighter through extended output.
- Multi-step reasoning chains with constraint satisfaction. Better at honoring "all of A, B, and C must hold" than the alternatives.
- Long-context refactoring tasks. The 200K window plus the structured-output reliability is a strong combo.
GPT-5.5 wins at
- Structured tool calls and function selection. Best-in-class JSON-mode reliability and function-calling accuracy.
- Complex agentic workflows with branching tool use. Picks the right tool more often, recovers from tool errors more gracefully.
- Tasks where the user expects "feel" of GPT-5 family. Hard to quantify, real for some customer segments.
Gemini 3.1 Pro wins at
- Anything with very long context (>250K tokens). The 2M window doesn't have an equivalent and the model holds attention across it surprisingly well.
- Vision-heavy workloads. Highest input resolution and best charts/figures parsing of the three.
- Cost-sensitive workloads. 60% discount vs Claude, 67% vs GPT-5.5 on input. At scale, this dwarfs other considerations.
The hidden margin: prompt caching
If you're going to remember one thing from this article, make it this: turn on prompt caching. All three providers offer it, and the savings are 80-90% on the cached portion. For most production workloads, this is the difference between a sustainable bill and a runaway one.
Anthropic's prompt caching is opt-in: you mark cache breakpoints in the API request and the system keeps the prefix for 5 minutes (default). Google's caching is similar. OpenAI uses an automatic prefix cache that just works if your requests share long prefixes.
The mental model: any prompt segment that's identical across 2+ requests in a 5-minute window should be cached. System prompts. Few-shot examples. RAG context that's stable for a session. Conversation history within a session. All of it.
What to track per customer (because you should)
The pricing change cadence in 2026 has been roughly every 6 weeks across the three providers. That's faster than most teams can update their internal cost models manually. The way out is to instrument per-customer LLM cost tracking and let the actual data drive your pricing decisions instead of your assumptions.
UsageBox does this with a thin wrapper around your model calls. Each call records token counts and the model name, and the dashboard shows you cost-per-customer in real time. When OpenAI cuts GPT-5.5 input prices next month, you'll see the impact on your margins the next day instead of catching it on the next invoice.
The honest recommendation, May 2026
Start every new workload on Gemini 3.1 Pro until proven otherwise. The price is half what the alternatives charge, the capability is in the same ballpark, and the 2M context window saves you from chunking work you'd otherwise have to do. Then run a quality A/B on representative workloads. If the quality gap to Claude or GPT-5.5 is large enough to justify 2x-3x the bill, switch. If not, you save real money.
For very-long-context work (>250K tokens) the choice is forced: Gemini is the only one with the window. For code-heavy and constraint-heavy work, Claude Opus 4.7 still has the edge. For tool-heavy agent loops, GPT-5.5 still has the edge.
For everything else, the price discount is bigger than the quality gap.
Related reading
- Gemini Pro Free Tier Killed (April 2026)
- Stripe AI Token Billing Without the Waitlist
- Claude Sonnet vs Kimi vs DeepSeek: Real Pricing
- Gemini API Free Tier (2026), the free-tier-after-billing gotcha for Gemini calls
- Claude API Billing in 2026, per-model limits + caching math for Opus/Sonnet/Haiku