DeepSeek vs ChatGPT: 17x Price Gap

A one-million-token context window is not a one-million-token price. Every major provider sells the window at one rate and then charges a different, higher rate once a prompt crosses a threshold — 200,000 tokens at Google and xAI, 272,000 at OpenAI. Above the line, the entire request bills at the higher rate, not just the excess tokens.

That distinction is where budgets break. A team that reads "1M context" and assumes a flat rate plans for $2.00 on a million-token call and receives a $4.00 invoice. Worse, the cliff is invisible in most cost calculators, because they multiply tokens by a single rate. This article maps every threshold in the dataset and works out when paying the surcharge is rational.

Long-context tiers are the least visible line item in an LLM bill.

Every threshold, October 2026

Model	Window	Tier starts	Input above tier	Output above tier
Gemini 3.8 / 3.7 Flash	1M	200K	$1.50	$7.50
Gemini 3.1 Pro	1M	200K	$4.00	$18.00
Gemini 2.5 Pro	1M	200K	$2.50	$15.00
Grok 4.3	1M	200K	$2.50	$5.00
Grok 4.6	500K	200K	$4.00	$12.00
GPT-6 Astra	1.05M	272K	$20.00	$75.00
GPT-6 Sol / 6.1 Sol	1.05M	272K	$4.00	$15.00
GPT-6 Luna	1.05M	272K	$0.20	$0.75
GPT-5.6 family	1.05M	272K	2× base	1.5× base
Claude Sonnet 5 / Opus 5 / Fable 5.1	1M	None	—	—
DeepSeek V4.1 Flash	1.1M	None	—	—
DeepSeek V4 Pro	1M	None	—	—

What a million-token prompt actually costs

One call, one million input tokens, no output. This is the number that matters when someone says "we'll just put the whole codebase in the prompt".

Model	Cost for 1M input	Effective rate / 1K	Tier applies?
DeepSeek V4.1 Flash	$0.30	$0.0003	No
Gemini 3.8 Flash	$1.50	$0.0015	Yes (2×)
Claude Sonnet 5	$2.00	$0.0020	No
Grok 4.3	$2.50	$0.0025	Yes (2×)
GPT-6 Sol	$4.00	$0.0040	Yes (2×)
Gemini 3.1 Pro	$4.00	$0.0040	Yes (2×)
Grok 4.6	$4.00 (capped at 500K)	—	Yes (2×)
GPT-6 Astra	$20.00	$0.0200	Yes (2×)

The spread is 67× between DeepSeek and GPT-6 Astra for the same nominal one-million-token window. Model choice dominates every optimisation technique at this length — compression included.

Why crossing the line is worse than it looks

The surcharge applies to the whole request. Take GPT-6 Sol: below 272K, input bills at $2.00 per million; above it, $4.00. A 400,000-token prompt therefore costs $1.60, not the $0.80 a naive multiplication suggests.

Now compress it to 250,000 tokens. You removed 37.5% of the tokens and saved 69% of the cost:

Prompt	Tokens	Rate	Input cost
Uncompressed	400,000	$4.00 (tier)	$1.60
Compressed	250,000	$2.00 (base)	$0.50
Saving	−37.5%	—	−69%

This is the highest-leverage optimisation in the entire field, and it is why "just add more context" is usually the most expensive way to improve an answer. The goal is not to fit the window; it is to stay below the threshold.

Three ways to compress without losing the answer

1. Rerank before you retrieve. Most RAG pipelines retrieve 50 documents and use all 50. A cross-encoder reranker that keeps the best 8 usually preserves answer quality while cutting tokens by 80%. The cost of the reranker is a fraction of the tokens it saves.

2. Summarise the middle, keep the edges. Models attend most reliably to the start and end of a long context. Compressing the middle third into a structured summary — entities, decisions, numbers — preserves most of the signal at a third of the length.

3. Extract, then reason. Instead of sending 300 pages and asking a question, run a cheap extraction pass with GPT-6 Luna that pulls only the fields the question needs, then send 2,000 tokens to the expensive model. Two calls, a fraction of the cost.

When paying the surcharge is correct

Compression has a failure mode: it removes the sentence that mattered. Four workloads justify the full window:

• Legal and contract review — the missing clause is the point of the review.
• Whole-repository analysis — cross-file dependencies cannot be summarised away.
• Long transcripts — a summary of a two-hour call loses exactly the details being audited.
• One-shot synthesis — where chunking would produce five inconsistent answers instead of one.

In those cases the comparison is not "long context versus cheap" but "long context versus a human reading it", and the surcharge is trivial.

The cheapest long-context model is not the obvious one

Claude Sonnet 5 has no surcharge and costs $2.00 for a million-token prompt — half of GPT-6 Sol's $4.00 despite identical sticker rates. DeepSeek V4.1 Flash costs $0.30 and offers 1.1M tokens flat, which is 13× cheaper than GPT-6 Sol at the same length.

If your workload genuinely needs very long prompts, the model shortlist is DeepSeek first on price, Claude second, and everything else a distant third. If your workload merely drifted above 272K by accident, the answer is compression, not a different provider.

FAQ

What is a long-context surcharge?

A higher per-token rate that applies to the entire request once the prompt exceeds a provider threshold — 200,000 tokens at Google and xAI, 272,000 at OpenAI. Anthropic and DeepSeek do not apply one.

Does the surcharge apply only to the extra tokens?

No. The whole request moves to the higher rate. A 300,000-token prompt on GPT-6 Sol bills all 300,000 tokens at $4.00, not 272,000 at $2.00 and 28,000 at $4.00.

Which model is cheapest for a 1M-token prompt?

DeepSeek V4.1 Flash at $0.30, followed by Gemini 3.8 Flash at $1.50 and Claude Sonnet 5 at $2.00.

Is a larger context window always better?

No. Above the threshold, a larger window costs twice as much per token, and accuracy on very long inputs is not guaranteed. Compression below the threshold is usually the cheaper and more accurate option.

Related reading

• Hidden Costs in LLM APIs
• LLM Cost Formula Explained
• The Cheapest Model Depends on Your Input

Check your own prompt length against every threshold in the free LLM cost calculator, or compare models directly in the full price ranking.