The LLM API Price War, Ranked: What Developers Are Actually Paying Now

We ran a 4M-token agent eval on GPT-5.5 last week. The bill: $124. We re-ran it on DeepSeek V4 Flash. Total cost: $0.97. The gap is real, the math checks out, and it fundamentally rewrites the unit economics of every agentic workload shipping in 2026.

The LLM pricing landscape has shifted twice this year alone. OpenAI cut o3 by 80% in January. DeepSeek slashed V4 Pro by 75% on April 27, 2026, plus cut cache-hit costs to one-tenth of launch price. Google’s Gemini 2.5 Flash-Lite is now $0.10 per million input tokens. This isn’t incremental. It’s the death of the premium-only model in production AI.

We haven’t tested DeepSeek V4 Pro at scale yet. We’ve run Flash on real agent jobs and didn’t catch a single output quality regression. Here’s what the numbers actually are today.

Price Tiers, Head-to-Head

The outliers? DeepSeek V4 Flash output is 107x cheaper per token than GPT-5.5. V4 Pro is 27x cheaper. For agent workloads where output tokens dominate the bill (agentic loops, long-horizon synthesis, tool-call chains), this isn’t a minor discount—it’s a business model reset.

Why o3 Matters (More Than You Think)

OpenAI’s 80% price cut on o3 didn’t come with a quality regression. Benchmark data on ARC Prize performance confirmed identical reasoning quality at the new price floor. That’s important: you’re not trading accuracy for pennies. You’re getting a better deal on the same inference engine.

At $2/$8 per million tokens, o3 sits between Gemini 3.1 Pro and the budget tier. It’s no longer the “premium reasoning only if you have budget” model. It’s competitive on cost and unmatched on reasoning depth.

The catch: you’ll burn tokens on long chains of thought. o3 is not a drop-in for GPT-5.5 on simple completions. Use it for the hard problems that need reasoning. Use Flash for the commodity stuff.

DeepSeek’s Promo Play (Expires May 31)

DeepSeek V4 Pro at 75% discount until May 31, 2026, sits at $0.27/$1.10 per million tokens. That’s a deliberate undercut of every Western model. The cache-hit pricing reset is the actual move here: input cache hits now cost 1/10th of the old rate, effective April 26. If you’re not caching agent context or RAG documents, you’re leaving money on the table.

We haven’t benchmarked V4 Pro against Claude Opus at production load yet. The spec looks competitive. The price is absurd. We’ll have a full eval by mid-June.

DeepSeek V4 Flash, at $0.14/$0.28 per million tokens, is the clearest win for token-heavy workloads. We’ve run three real agent projects on Flash. Output quality held steady. Latency was acceptable. Billing came in at roughly what we predicted. No surprises is the highest compliment we can give a model at that price.

Google’s Sub-$1 Floor: Gemini 2.5 Flash-Lite

Gemini 2.5 Flash-Lite at $0.10/$0.40 per million tokens is Google’s answer to “what if we price by actual cost?” The model is fast, the context is competent, and the price point forces every other player to justify premium positioning.

For chatbots, documentation search, and single-turn synthesis, Flash-Lite is a no-brainer. We’ve deployed it for customer-support automation and it’s solid. For agentic work or multi-turn reasoning, you’ll want more horsepower. But for a tier that costs less than a tenth of GPT-5.5, Google’s shipping real value here.

What Wins in Each Bracket

Chatbots & simple completion: Gemini 2.5 Flash-Lite or DeepSeek V4 Flash. Both are fast, both are cheap, both ship decent output. Slight edge to Flash-Lite on speed, Flash on output quality across non-English. Pick one and move on.

RAG & document retrieval: DeepSeek V4 Pro with cache hits. The cache-cost reset is what actually changes your bill. Cache a 50k-token knowledge base once, hit it 100 times. You pay ~$0.014 on the first request (50k tokens × $0.27/M), then $0.0014 on each subsequent cache hit (1/10th rate = $0.027/M × 50k tokens). Claude Opus caching runs the same 1/10th multiplier, but its cache-hit rate ($0.50/M) is ~18x higher than V4 Pro’s ($0.027/M). That gap compounds across a live RAG system.

Agent reasoning & hard problems: o3 still wins here. The 80% cut makes it justifiable on cost. Claude Opus 4.6 at $5/$25 is close, and Anthropic’s caching is solid, but o3’s reasoning chops are unmatched in the mid-tier price bracket.

Token-intensive synthesis (long-form writing, code generation, SQL from natural language): DeepSeek V4 Flash. The output-token cost is so low that marginal quality differences don’t matter. You’re paying roughly $0.003 for a 10k-token response on Flash ($0.28/M × 10k). The same output on GPT-5.5 costs $0.30. Flash won’t beat GPT-5.5 on prose quality, but it’s good enough for 99% of production workloads, and the economics are undeniable.

The Number That Actually Matters: Output Tokens Dominate

This is the footnote that kills most cost models. In agent workloads—tool calls, function returns, reasoning traces, multi-step synthesis—output tokens are 70–90% of your bill.

A typical agent loop: you send a 200-token user query + 2k-token system prompt + 5k-token context (input cost: ~$0.001 on DeepSeek Flash at $0.14/M). The model thinks, calls a tool, gets a result, thinks again, and returns a 3k-token response. The output cost: ~$0.001 at $0.28/M. You’re not paying for input; you’re paying for the thinking.

This shifts which model is actually cheapest. GPT-5.5 and Claude Opus are close on input price ($5/M each). On output, GPT-5.5 is 4x more expensive than o3 ($30 vs $8) and 107x more expensive than Flash ($30 vs $0.28). In a 10-round agent interaction, that gap compounds hard. Switch to Flash and the savings are not marginal.

Cache hits are the second-order effect. If your agent system prompt and context are stable across calls, caching cuts your effective cost per interaction by 70–80%. DeepSeek’s cache-hit price is so low now that it’s worth redesigning your prompt strategy around cached context.

Where We’d Start a Project Today

If we were spinning up a new agent system in May 2026, we’d tier it: Gemini 2.5 Flash-Lite for the MVP (cheap, fast, zero regret if we pivot). DeepSeek V4 Flash for production once we’ve validated the use case (quality is solid, cost is undeniable). o3 for the reasoning problems that actually need it (benchmarked performance advantage is real).

We’re not giving GPT-5.5 or Claude Opus Tier 1 status here — not because they’re bad, they’re excellent — but because the value prop has evaporated. You’re paying premium prices for marginal quality gains in most production scenarios. We’ll re-benchmark in 60 days because this pricing war is not done. But today, the budget models are shipping real output, and the premium models have to justify themselves on actual reasoning lift, not just brand.

Internal References

• Learn more about Gemini API Pro free tier changes
• See GitHub Copilot’s new token billing in June 2026
• Understand Replit Agent 3 effort pricing