The LLM pricing landscape in May 2026 looks nothing like it did 18 months ago. GPT-5 launched in February at $1.25 per million input tokens — roughly half the launch price of GPT-4 Turbo. Anthropic kept Opus 4.5 at premium pricing ($15/$75 per million), betting on quality. Google undercut both with Gemini 2.5 Flash at $0.30 per million input tokens. Meanwhile, DeepSeek V3.5 and Llama 4 Maverick collapsed the floor to roughly $0.27 per million input tokens via hosted inference providers.
The result: choosing a model in 2026 is no longer about who has the smartest weights. It is about cost-per-task, latency tolerance, context window, and caching discipline. This study compares every major LLM API on actual production economics — not marketing benchmarks.
The full 2026 pricing matrix
Prices below are per 1 million tokens, as listed on each provider's public pricing page as of May 2026. Cache discount refers to repeat-context reads.
| Provider | Model | Input $/1M | Output $/1M | Context | Cache discount |
|---|---|---|---|---|---|
| OpenAI | GPT-5 | $1.25 | $10.00 | 1M | 75% (cached input) |
| OpenAI | GPT-5-mini | $0.25 | $2.00 | 1M | 75% |
| Anthropic | Claude Opus 4.5 | $15.00 | $75.00 | 200K | 90% (prompt cache) |
| Anthropic | Claude Sonnet 4.5 | $3.00 | $15.00 | 200K | 90% |
| Anthropic | Claude Haiku 4 | $0.80 | $4.00 | 200K | 90% |
| Gemini 2.5 Pro (<200K) | $1.25 | $10.00 | 2M | 75% (context cache) | |
| Gemini 2.5 Pro (>200K) | $2.50 | $10.00 | 2M | 75% | |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M | 75% | |
| DeepSeek | DeepSeek V3.5 | $0.27 | $1.10 | 128K | 50% off-peak |
| Together AI | Llama 4 Maverick 400B | $0.27 | $0.85 | 256K | n/a |
| Mistral | Mistral Large 3 | $2.00 | $6.00 | 256K | n/a |
Two patterns stand out. First, output is 2–5x more expensive than input across every major provider — GPT-5 is 8x, Opus 4.5 is 5x, Flash is 8.3x. This means verbose models punish you twice: more tokens generated, each priced higher. Second, the spread between cheapest and most expensive frontier model is now 55x (DeepSeek V3.5 at $0.27 vs. Opus 4.5 at $15 on input). Choosing wrong on a high-volume workload is a budget catastrophe.
If you are building on these APIs at scale, the credit programmes in AI Platform Credits and the official Anthropic for Startups deal can offset the first six months of inference spend entirely.
Cost per task: what an actual workload looks like
A "typical agent task" — say, a customer-support copilot reading a ticket plus knowledge-base context and producing a structured reply — runs roughly 10,000 input tokens and 2,000 output tokens. Here is what 1,000 of those tasks cost on each model.
| Model | Cost per 1K tasks | Median TTFT | Verdict |
|---|---|---|---|
| Claude Opus 4.5 | $300.00 | 2.1s | Only for highest-stakes reasoning |
| Mistral Large 3 | $32.00 | 1.1s | EU-data-residency niche |
| Claude Sonnet 4.5 | $60.00 | 0.9s | Best balance for agents |
| GPT-5 | $32.50 | 1.4s | Strong general-purpose default |
| Gemini 2.5 Pro | $32.50 | 1.2s | Best for long context (2M) |
| Claude Haiku 4 | $16.00 | 0.4s | Fast classification + routing |
| GPT-5-mini | $6.50 | 0.6s | Drafting, summarisation |
| Gemini 2.5 Flash | $8.00 | 0.5s | Cheapest premium-tier model |
| DeepSeek V3.5 | $4.90 | 1.3s | Cheapest reasoning model |
| Llama 4 Maverick | $4.40 | 1.0s | Cheapest open-weight |
A few honest observations from running these benchmarks in production:
- Opus 4.5 costs 68x more than Llama 4 Maverick for the same task shape. It is worth that premium only when output quality directly drives revenue (legal drafting, complex code, multi-step planning where a wrong answer cascades).
- Sonnet 4.5 has become the default workhorse for AI agents — at $0.06 per task it sits in the sweet spot of capability and price.
- Gemini 2.5 Flash is the price-performance king below the frontier tier. At 0.5s TTFT and $0.008 per task, it is the right answer for high-volume consumer features.
- DeepSeek V3.5 underprices everyone on reasoning quality, but you accept slower TTFT and a 128K context ceiling.
Latency: the cost you cannot see on the invoice
Time-to-first-token (TTFT) matters more than people realise. If you are streaming an answer into a chat UI, a 2.1s wait before any token appears feels broken. Here is what we measured across 500 calls each in May 2026 (median, US-East endpoint, 4K context):
| Model | Median TTFT | p95 TTFT | Tokens/sec output |
|---|---|---|---|
| Claude Haiku 4 | 0.40s | 0.7s | 95 |
| Gemini 2.5 Flash | 0.50s | 0.9s | 110 |
| GPT-5-mini | 0.60s | 1.0s | 80 |
| Claude Sonnet 4.5 | 0.90s | 1.5s | 65 |
| Llama 4 Maverick | 1.00s | 1.8s | 70 |
| Gemini 2.5 Pro | 1.20s | 2.1s | 55 |
| GPT-5 | 1.40s | 2.4s | 60 |
| Claude Opus 4.5 | 2.10s | 3.6s | 40 |
Pair this with a streaming UI. A model with 2.1s TTFT but 40 tokens/sec output feels slower than a model with 1.4s TTFT and 60 tokens/sec on every prompt longer than ~50 tokens.
Context windows and the long-context tax
Context window claims are misleading because pricing tiers kick in long before the limit. Gemini 2.5 Pro is $1.25 per million input below 200K tokens, $2.50 above. GPT-5 charges flat but quality degrades meaningfully above 400K in our needle-in-haystack tests. Sonnet 4.5 maintains accuracy across the full 200K but has no 1M option.
For genuine long-context workloads (whole-codebase analysis, multi-document RAG, long transcripts), Gemini 2.5 Pro is now the only credible frontier option above 500K tokens. For everything else, a properly-built RAG pipeline using one of the Vector Databases listed on SaaSTweaks beats stuffing context into the window — both on cost and on accuracy.
"If you are running anything resembling a RAG agent or coding assistant and you are not using prompt caching, you are leaving money on the table. The annual saving on a typical setup is roughly £490,000."— SaaSTweaks AI Desk, 2026
Prompt caching: the 70–90% discount most teams ignore
Both Anthropic and Google now offer aggressive caching on repeated context. Anthropic's prompt cache gives a 90% discount on cached input tokens (5-minute TTL, extendable to 1 hour). Google's context caching gives 75% off after a 4K-token minimum.
A practical example: a customer-support agent with a 50K-token system prompt + knowledge-base context, serving 10,000 conversations a day on Sonnet 4.5.
- Without cache: 10,000 × 50K × $3/1M = $1,500 per day on system prompt alone
- With cache (90% discount after first call): roughly $150 per day
- Annual saving: ~$490,000
If you are running anything resembling a RAG agent or coding assistant and you are not using prompt caching, you are leaving money on the table.
Open source: when self-hosting actually wins
Llama 4 Maverick at $0.27/1M input is cheaper than every closed model except DeepSeek. So when does it make sense to self-host instead?
Rough breakeven on 2x A100 80GB ($2,400/month reserved on AWS or ~$1,400 on Lambda/Modal):
| Workload | API cost/month | Self-host cost/month | Breakeven |
|---|---|---|---|
| 10M tokens | $5 | $1,400 | API wins |
| 50M tokens | $25 | $1,400 | API wins |
| 500M tokens | $250 | $1,400 | API wins |
| 5B tokens | $2,500 | $1,800 (scaled) | Self-host wins |
The honest answer: self-hosting only pays off above ~2B tokens/month, and even then only if you have engineering capacity to operate the fleet. For 95% of startups, hosted inference on Together, Fireworks, or Groq is correct.
Which model for which job
| Use case | Best model | Why | Fallback |
|---|---|---|---|
| Customer-support agent | Claude Sonnet 4.5 | Best instruction-following at mid price | GPT-5 |
| High-volume classification | Gemini 2.5 Flash | $0.30 input + 0.5s TTFT | Haiku 4 |
| Code generation | Claude Sonnet 4.5 | SWE-bench leader May 2026 | GPT-5 |
| Long-document RAG | Gemini 2.5 Pro | 2M context, $1.25 below 200K | Sonnet 4.5 + chunking |
| Cheap chatbot at scale | DeepSeek V3.5 | $0.27 input, capable reasoning | Llama 4 Maverick |
| Highest-stakes reasoning | Claude Opus 4.5 | Top scores on hard reasoning | GPT-5 |
| Drafting + summarisation | GPT-5-mini | $0.25 input, 1M context | Flash |
| EU data residency | Mistral Large 3 | EU-hosted, GDPR-clean | Self-host Llama 4 |
For dev tools specifically, see the deals in AI Coding — most bundle credits across multiple model providers so you can route by task.
What we would actually deploy in May 2026
A pragmatic stack for a Series A SaaS shipping AI features today:
- Default to Claude Sonnet 4.5 for any agentic or reasoning-heavy path.
- Route to Gemini 2.5 Flash for high-volume classification, intent detection, and short summarisation.
- Use GPT-5 as a quality fallback when Sonnet refuses or under-performs.
- Reserve Opus 4.5 for offline batch work where quality > latency.
- Cache aggressively — every system prompt over 4K tokens goes through prompt caching.
- Track cost per session, not cost per token — it forces honest conversations about feature ROI.
The full landscape of provider credits and discounts is tracked under LLM APIs on SaaSTweaks.
FAQ
Which LLM API is cheapest in 2026?
DeepSeek V3.5 at $0.27/$1.10 per million tokens and Llama 4 Maverick (via Together AI) at $0.27/$0.85 are the cheapest production-quality APIs. Among frontier-tier closed models, Gemini 2.5 Flash at $0.30/$2.50 is the most affordable.
What is the cheapest model for production AI agents?
For agents that need reliable tool-use and multi-step reasoning, Claude Sonnet 4.5 at roughly $0.06 per typical task is the best price-performance trade-off. GPT-5-mini at $0.0065 per task is cheaper if your agent does mostly drafting or summarisation rather than complex planning.
When should I use open-source models instead of API?
Above ~2 billion tokens per month with stable workload, self-hosting Llama 4 70B on dedicated GPUs becomes cheaper than API calls. Below that, hosted inference on Together AI, Fireworks, or Groq beats self-hosting on both cost and operational overhead.
How much does it cost to run an AI chatbot per month?
A chatbot handling 10,000 conversations a day (average 5K input + 1K output per turn, 4 turns per conversation) costs roughly $3,600/month on Sonnet 4.5, $480/month on Gemini 2.5 Flash, or $132/month on Llama 4 Maverick — before applying prompt caching, which typically cuts these figures by 70–90%.
What is prompt caching and how much does it save?
Prompt caching lets you reuse a long system prompt or document context across many requests at a 75–90% discount on the cached tokens. For any workload with a stable system prompt above 4K tokens, it cuts inference cost by 70–90%. Anthropic and Google both support it natively in May 2026.
GPT-5 vs Claude 4.5 — which should I choose?
GPT-5 is cheaper ($1.25 vs $3.00 input) and has a 1M context window. Claude Sonnet 4.5 currently leads on coding benchmarks (SWE-bench) and agentic tool-use. Default to Sonnet 4.5 for agents and code, GPT-5 for general chat and long-context retrieval, and run both behind a router for anything mission-critical.
