GPT-5.6 vs Gemini 3.7 Flash vs Grok 4.6: API Pricing Compared (August 2026)

“How much does the API actually cost” is a question every team building an AI feature ends up answering the hard way, after the first month’s invoice arrives. The honest answer changes constantly, current model families ship multiple pricing tiers, run time-limited introductory discounts, and charge differently depending on context length, so a number you saw quoted three months ago is probably already wrong. Here’s what OpenAI, Google, and xAI are actually charging as of this week, tier by tier.

The headline numbers, per million tokens

Model	Input	Output	Notes
GPT-5.6 Sol (OpenAI, flagship)	$5.00	$30.00	Effective July 30, 2026
GPT-5.6 Terra (OpenAI, balanced)	$2.00	$12.00	Effective July 30, 2026
GPT-5.6 Luna (OpenAI, high-volume)	$0.20	$1.20	Effective July 30, 2026
Gemini 3.7 Flash (Google)	$0.75	$3.75	Introductory, through Dec 31, 2026
Gemini 3.7 Flash (Google, standard)	$1.50	$7.50	Takes effect Jan 1, 2027
Grok 4.6 (xAI, under 200K context)	$2.00	$6.00	$0.50/M for cached input
Grok 4.6 (xAI, over 200K context)	$4.00	$12.00	Entire request billed at this rate
DeepSeek V4-Flash (off-peak)	$0.22	$0.66	01:00-04:00, 06:00-10:00 UTC
DeepSeek V4-Flash (peak)	$0.44	$1.32	All other hours

Every one of these tiers is priced for a different job, and the sticker price alone tells you less than the ratio between input and output cost. OpenAI keeps a fixed 6x multiplier across its entire GPT-5.6 family, which makes back-of-envelope math easy: estimate your input and output token counts, multiply by the Terra or Luna rate depending on the quality bar you need, and you have a real number. Google and xAI don’t hold that ratio as tightly, Gemini 3.7 Flash runs 5x and Grok 4.6 runs 3x, which matters more for output-heavy workloads like long-form generation than for short-answer or classification tasks.

The two details that actually move your bill

Two things in that table matter more than the headline number for a real production workload.

Time and context sensitivity aren’t edge cases. DeepSeek’s peak/off-peak split doubles your cost depending purely on when a request lands, and Grok 4.6’s 200K-token threshold doubles the entire request’s cost, not just the tokens past the line, the moment your context window crosses it. A workload that regularly sends long documents, large codebases, or extended conversation history as context needs to budget for the higher Grok tier by default, not treat it as an occasional edge case.

Introductory pricing has an expiration date. Gemini 3.7 Flash’s current rate is exactly half of what it becomes on January 1, 2027. If you’re building cost projections for a product launching or scaling into next year, model the standard rate, not the launch discount, or your projected margins will be wrong by a predictable amount on a predictable date.

Where prompt caching changes the calculus

None of the numbers above account for caching, and for most production agent or RAG setups, caching is where the real savings live. A system prompt, a set of few-shot examples, or a large retrieved-context block that repeats across calls is exactly the pattern caching is built for: OpenAI prices a cache read at a tenth of the standard input rate, and DeepSeek prices a cache hit on V4-Flash at roughly 50 times cheaper than a cache miss. A workload with a large, mostly-static prompt prefix can end up cheaper on a provider with a “more expensive” headline rate but stronger caching discounts than on one with a lower sticker price and a weaker cache. If you’re optimizing an existing LLM API bill rather than picking a provider from scratch, checking your actual cache hit rate before switching models is usually the higher-leverage move.

How to actually pick

Estimate your real input and output token counts from production traffic, not a demo, then run that ratio against each tier in the table above rather than comparing headline input prices alone, since output tokens cost several times more than input tokens across every provider here. If your workload sends long context regularly, price it against Grok’s over-200K rate and DeepSeek’s peak rate, not the best case. And if quality at the low end matters for your use case, don’t assume the cheapest tier is a like-for-like swap for a flagship model. Run your own eval set against the cheaper tier before committing a production workload to it. Our team runs exactly this kind of cost-versus-quality evaluation for clients migrating an AI feature between providers, and the answer is rarely just “pick the lowest number in the pricing table.”