LLM API Price Watch — June 2026 (OpenAI, Anthropic, Google)
Last updated 2026-08-08
A data-driven snapshot of OpenAI, Anthropic, and Google API pricing this month — with the blended cost per task and the exact point where local wins.
By Mohamed Meguedmi · 9 min read
Key Takeaways
- Google Gemini 2.5 Flash is the cheapest capable workhorse at $0.30 / $2.50 per 1M tokens. Only GPT-5 nano and GPT-5 mini undercut it, and both trade away reasoning quality to do so.
- Anthropic remains the premium tier. Claude Opus 4.7 holds at $15 / $75 per 1M — roughly 12× the input price of Gemini 2.5 Pro. You are paying for reasoning depth, not tokens.
- Cached input is the real lever. OpenAI and Anthropic both price cached reads at a ~10× discount, which turns a long, static system prompt from a recurring tax into a rounding error.
- Local break-even is tier-dependent: ~3M tokens/month if you would otherwise run Opus, but ~100M tokens/month versus Gemini 2.5 Flash. Volume, not vibes, decides.
- These numbers move almost monthly — always confirm against each vendor's live pricing page before you lock a budget.
Cloud LLM pricing in mid-2026 has stopped falling in a straight line. The headline flagships have converged near $1.25 input / $10 output per million tokens, the cheap tiers are in a race to the bottom under $0.10 input, and the premium reasoning models refuse to budge. This is our June 2026 board — the numbers we track, the ones that actually move a budget, and the point at which running the model yourself is the rational call.
The June 2026 price board at a glance
All figures below are list prices in USD per 1M tokens for standard (non-batch) synchronous API calls, current as of June 2026. Cached-input columns reflect the discounted read rate for prompt content served from provider-side cache.
| Provider | Model | Input $/1M | Output $/1M | Cached input $/1M | Context |
|---|---|---|---|---|---|
| OpenAI | GPT-5 | 1.25 | 10.00 | 0.125 | 400K |
| OpenAI | GPT-5 mini | 0.25 | 2.00 | 0.025 | 400K |
| OpenAI | GPT-5 nano | 0.05 | 0.40 | 0.005 | 400K |
| Anthropic | Claude Opus 4.7 | 15.00 | 75.00 | 1.50 | 200K |
| Anthropic | Claude Sonnet 4.6 | 3.00 | 15.00 | 0.30 | 1M |
| Anthropic | Claude Haiku 4.5 | 1.00 | 5.00 | 0.10 | 200K |
| Gemini 2.5 Pro | 1.25 | 10.00 | 0.31 | 1M | |
| Gemini 2.5 Flash | 0.30 | 2.50 | 0.075 | 1M | |
| Gemini 2.5 Flash-Lite | 0.10 | 0.40 | 0.025 | 1M |
Verify these against the sources directly: OpenAI API pricing, Anthropic pricing, and the Gemini API pricing docs. Google still applies a higher tier to Gemini 2.5 Pro requests above 200K input tokens ($2.50 / $15), so long-context Pro calls cost more than the table implies.
What changed since Q1 2026
Three moves define the quarter. First, OpenAI's mid tier got aggressive: GPT-5 mini at $0.25 / $2.00 is now the cheapest model that still handles multi-step tool use reliably, which pulls a lot of production traffic away from the flagship. Second, Google held Gemini 2.5 Flash flat but widened its lead on context — a full 1M window at $0.30 input is unmatched at that price. Third, Anthropic did not cut Opus. That is a deliberate signal: Opus 4.7 is positioned as a reasoning instrument for high-stakes work, not a volume commodity, and the price tells you exactly where they think it belongs.
The convergence is real but shallow. Four different models now sit at or near $1.25 / $10, yet their output quality on agentic and long-horizon tasks is anything but identical. Price parity is not capability parity.
Real cost per task — the blended view
List prices per million tokens are hard to feel. So we model a representative retrieval-augmented task: 8,000 input tokens (system prompt + retrieved context + question) and 2,000 output tokens (a structured answer). This 4:1 input-to-output ratio is typical of RAG, support, and coding-assist workloads. No caching applied.
| Model | Cost / task (8K in + 2K out) | Cost / 10K tasks | Relative |
|---|---|---|---|
| GPT-5 mini | $0.0060 | $60 | 1.0× |
| Gemini 2.5 Flash | $0.0074 | $74 | 1.2× |
| Claude Haiku 4.5 | $0.0180 | $180 | 3.0× |
| GPT-5 | $0.0300 | $300 | 5.0× |
| Gemini 2.5 Pro | $0.0300 | $300 | 5.0× |
| Claude Sonnet 4.6 | $0.0540 | $540 | 9.0× |
| Claude Opus 4.7 | $0.2700 | $2,700 | 45× |
Two things jump out. GPT-5 mini edges out Gemini 2.5 Flash on this exact mix purely because its output rate is $2.00 versus $2.50 — flip the ratio toward more output and Flash regains the lead, so profile your own traffic. And Opus 4.7 costs 45× the cheapest option. That multiplier is only worth paying when a wrong answer is expensive: legal review, complex refactors, financial reasoning. For classification, extraction, or first-draft generation, it is indefensible. Run your own mix through our cost calculator before you standardize on a model.
Cached input and batch — the discounts that actually move the bill
If you send the same 6,000-token system prompt on every request, you are paying for it every time — unless you cache. In June 2026 the cached-read discount is roughly 10× on OpenAI and Anthropic, and about 4× on Gemini 2.5 Pro. On a workload with a large static preamble, this is the single biggest lever available.
Worked example: a 6,000-token static system prompt served to 100,000 requests.
- GPT-5, uncached: 6,000 × 100,000 × $1.25 / 1M = $750 in system-prompt input alone.
- GPT-5, cached reads: the same content at $0.125 / 1M ≈ $75.
A $675 saving on one prompt component, for the cost of structuring your calls so the static prefix is cache-eligible. Layer batch processing on top — both OpenAI and Anthropic offer ~50% off asynchronous batch jobs — and non-latency-sensitive pipelines (nightly summarization, backfills, evals) roughly halve again. The blended board price is the ceiling, not the number you should actually pay.
When does local beat the API?
The honest answer is: it depends entirely on which tier you are replacing. A local Qwen3-Coder 30B or Llama-class model competes with the mid tier, not with Opus-grade reasoning, so compare like for like. We model a single 24GB-class GPU amortized over 24 months plus power: call it ~$105/month all-in. We blend API cost at a 70% input / 30% output ratio.
| API tier replaced | Blended $/1M (70/30) | Local cost/mo | Break-even volume |
|---|---|---|---|
| Gemini 2.5 Flash | $1.06 | ~$105 | ~99M tok/mo |
| GPT-5 | $3.88 | ~$105 | ~27M tok/mo |
| Claude Sonnet 4.6 | $6.60 | ~$105 | ~16M tok/mo |
| Claude Opus 4.7 | $33.00 | ~$105 | ~3.2M tok/mo |
The takeaway: if your workload only ever needed Flash-tier quality, the API is almost impossible to beat — you would need to push ~100M tokens a month before local hardware pays for itself, and that ignores your engineering time. But if you have been reaching for Opus or Sonnet and your quality bar can be met by a well-quantized open model such as Qwen3-Coder 30B Q4_K_M, break-even arrives fast — a few million tokens a month. See the Qwen3-Coder model card and the Ollama qwen3 page for quantization options, and cross-reference quality on our benchmarks before assuming parity.
Verdict — which API to default to
There is no single winner, but there are clear defaults by job:
- High-volume, cost-sensitive (classification, extraction, drafts): Gemini 2.5 Flash. The 1M context at $0.30 input is the best value on the board, and GPT-5 mini is a near-tie for output-light traffic.
- Balanced production agent: GPT-5 or Gemini 2.5 Pro at $1.25 / $10. Pick on ecosystem and tool-use behavior, not price — they are identical on cost.
- Hardest reasoning, correctness-critical: Claude Opus 4.7. Expensive by design; reserve it for the tasks where a wrong answer costs more than $0.27.
- Predictable heavy internal usage above the break-even lines: self-host a Qwen3 or Llama-class model.
For programmatic comparisons, the full price and benchmark dataset behind this article is available through the BestLLMfor public API (CC BY 4.0) and the open-source MCP server, so you can wire live pricing into your own dashboards. Browse every model on the catalog.
| Use case | Pick | Why |
|---|---|---|
| Cheapest capable | Gemini 2.5 Flash | $0.30/$2.50, 1M context |
| Balanced flagship | GPT-5 / Gemini 2.5 Pro | $1.25/$10, strong tool use |
| Top reasoning | Claude Opus 4.7 | Quality worth the 45× premium — sometimes |
| Long static prompts | Any with cache | ~10× cached-read discount |
| Heavy internal volume | Local Qwen3/Llama | Break-even from ~3M tok/mo vs Opus |
Frequently asked questions
Which LLM API is cheapest in June 2026?
For most real workloads, Google Gemini 2.5 Flash at $0.30 input / $2.50 output per 1M tokens offers the best value at usable quality. GPT-5 nano ($0.05 / $0.40) and GPT-5 mini ($0.25 / $2.00) are cheaper still on paper, but nano is limited to simple tasks and mini trades away some reasoning reliability. Profile your input/output ratio — for output-heavy work Flash usually wins, for input-heavy work mini can edge ahead.
Is Claude Opus 4.7 worth 45× the cost of GPT-5 mini?
Only when correctness has real dollar value. On our representative task Opus costs $0.27 versus $0.006 for GPT-5 mini. For legal analysis, complex code refactoring, or financial reasoning where a wrong answer is expensive, the premium is justified. For classification, extraction, or first drafts, it is not — route those to a cheaper tier.
Do cached-input discounts really matter?
Enormously, if you reuse a large static prompt. The cached-read rate is roughly 10× cheaper on OpenAI and Anthropic. A 6,000-token system prompt sent to 100,000 requests costs about $750 uncached on GPT-5 versus roughly $75 cached — a $675 saving on one prompt component. Structure your calls so the static prefix is cache-eligible.
When should I self-host instead of using an API?
It depends on the tier you are replacing. Versus Gemini 2.5 Flash, a single 24GB-class GPU only breaks even near 100M tokens/month. Versus Claude Opus 4.7, break-even can arrive by ~3M tokens/month. If a well-quantized open model like Qwen3-Coder 30B meets your quality bar and you have steady heavy volume, local pays off quickly — but confirm quality on independent benchmarks first.
How often do these prices change?
Frequently — meaningful moves happen almost every month, and mid-tier models are especially volatile. Treat any published price table, including this one, as a snapshot. Always confirm against each vendor's live pricing page before committing a production budget.
Methodology note: all task-cost figures are computed from June 2026 list prices at the stated token ratios, exclude caching and batch discounts unless noted, and assume USD standard-tier synchronous pricing. See our benchmarks for the quality data that should accompany any cost decision.
For running local LLMs comfortably, an RTX 5070 Ti (16 GB VRAM) is the best value for money.
Amazon Check RTX 5070 Ti price →As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.