LLM API Pricing Compared in 2026: GPT-6 Astra vs Claude Opus 5.5 vs Gemini 4 vs DeepSeek

The same million input tokens and million output tokens cost $60 on OpenAI's GPT-6 Astra, $24 on Claude Opus 5.5, $12 on Gemini 4 Carbon through November 30, and $5.28 on DeepSeek V4-Pro at peak rates — an 11x spread between four frontier models that all advertise million-token context windows, and 23x if you catch DeepSeek's off-peak meter. November 2026 is the worst month in years to lock a rate card without checking the expiration dates. Three clocks are running at once: OpenAI's GPT-5.6 Sol promotional pricing holds "at least through November 21," Google doubles Gemini 4 Carbon and Argon from $2/$10 to $4/$20 on December 1, and the Gemini Flash line doubles on January 1, 2027. Anthropic spent 2026 pushing the other way — Opus 5.5 shipped at $4/$20, down from Opus 5's $5/$25, and a scheduled Sonnet increase was cancelled outright. This guide breaks down LLM API pricing across the four vendors on numbers pulled from their own pricing pages this week: list rates, cache economics, long-context surcharges, and worked bills at real volume.

What Changed in 2026: Two Vendors Raised Prices While Two Cut Them

April 2026: OpenAI ships GPT-5.5 at $5 per million input and $30 per million output — roughly double GPT-5.4 — and positions it as the agent-and-coding flagship. By November the menu has moved again. The current OpenAI rate card runs four text models: GPT-6 Astra at $10/$50, GPT-5.6 Sol at $4/$20 (a promotional rate the pricing page says holds "at least through November 21"), GPT-5.6 Terra at $2/$12, and GPT-5.6 Luna at $0.20/$1.20. Every model carries a 1.05 million-token context window with 128K maximum output, and every published price is a short-context price: beyond 272,000 input tokens, OpenAI bills input at 2x, cached input at 2x, and output at 1.5x. Filling Astra's window is not a $10-per-million exercise; it is a $20 one.

Anthropic's 2026 went the opposite way. Claude Opus 5.5 landed at $4/$20 — a fifth cheaper on input than the $5/$25 Opus 5 it succeeded. The pricing docs carry an unusual footnote: Sonnet 5's $2/$10 was announced as introductory pricing through August 31, 2026, the planned rise to $3/$15 on September 1 never happened, and $2/$10 is now the standard rate. The budget tier got its own reset — Haiku 5.5 costs $0.10/$0.50 for prompts up to 100,000 tokens, one-fifth of Haiku 4.5's $1/$5, though prompts beyond that threshold jump to $0.50/$2.50.

Google's Gemini 4 family (Carbon, Argon, Hydrogen) launched on intro pricing that doubles in December: Carbon and Argon move from $2/$10 to $4/$20 on December 1, and cache reads rise from $0.10 to $0.20 the same day. The Flash line gets its own doubling on January 1, 2027 — 3.8 Flash goes from $0.75/$3.75 to $1.50/$7.50. Gemini 2.5 Pro is retired; requests route to 3.8 Flash at Flash rates from December 1.

DeepSeek restructured its meter instead of its levels. The V4 generation splits into deepseek-flash and deepseek-v4-pro, both with 1M context and up to 384K output, billed at peak and off-peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays — seven hours a day, five days a week. Every other hour, including the whole weekend, bills at exactly half.

LLM API Pricing at a Glance: Four Frontier Rate Cards

Read the cache column as a multiple of list input and the gaps widen: Astra charges 10 percent of its input rate for a cache read, Opus 5.5 charges 5 percent, Carbon 5 percent until it doubles, and DeepSeek about 3 percent with no write premium. Batch halves everything on the first three — Astra's input drops to $5, Opus 5.5 to $2, Carbon to $1 — while DeepSeek's batch equivalent is simply the off-peak clock that covers roughly 79 percent of the week.

The Frontier Tier: What $60, $24, $12, and $5.28 Buy

Price the same benchmark — 1M tokens in, 1M tokens out — and the frontier splits into three bands. Astra is alone at the top: $60 a blended million, a knowledge cutoff of April 30, 2026, and the only model here whose output cap (128K) is a fraction of its rivals' decode ranges. Sol, Opus 5.5, and Carbon-in-December cluster at $24, which is the number worth remembering: Google's intro price is a loan, and on December 1 Carbon lands exactly on the Sol/Opus line. V4-Pro undercuts all of it at $5.28 peak, $2.64 off-peak — with two strings attached. The concurrency cap is 500 requests, versus effectively pooled capacity at the hyperscalers, and DeepSeek's published web-tool add-ons ($0.75 per 1,000 searches or fetches, $0.225 per 1,000 pages parsed) are separate line items rather than bundled features.

Anthropic also sells a tier above Opus now: Claude Fable 5.2 at $8/$40, with the limited-availability Mythos 5.2 alongside. If your workload genuinely needs it, the price is real; most teams evaluating frontier models will never touch it.

The Workhorse Tier: Where Real Volume Lives

Frontier flags get planted in demos; bills get written in the workhorse tier. Same four vendors, same structure:

The budget shelf below that tier is the deepest it has ever been. OpenAI's Luna prices output at $1.20 — the same number as deepseek-flash's peak rate — with cached input at $0.02. Haiku 5.5 undercuts everything at $0.10/$0.50 for sub-100K prompts. Gemini 3.5 Flash-Lite holds at $0.30/$2.50 with no announced increase. And GPT-5.1, the previous OpenAI workhorse at $1.25/$10, is still on sale; price trackers that snapshot daily have recorded no move since August. None of these numbers requires a negotiated contract, a reserved-commitment discount, or a sales call — they are the metered list prices any credit card can hit today.

Prompt Caching Is the Third Price Column

Every vendor now prices a cached token differently from a fresh one, and the deltas are large enough to reorder a comparison. OpenAI's structure on the 5.6 generation mirrors Anthropic's: cache reads run at 10 percent of list input (Sol $0.40, Astra $1.00), but writes carry a premium — $5.00 per million for a 30-minute TTL and $10.00 for a 12-hour one on Sol. Anthropic charges 1.25x list input for a 5-minute write and 2x for an hour, then reads back at 5 percent on current models — $0.10 on Sonnet 5.5, $0.20 on Opus 5.5 — with break-even after a single read. Google only supports implicit caching on Gemini 4, which means you cannot pin a retrieval corpus the way explicit caching APIs let you — a real constraint for RAG stacks, which we compared in our vector database review — and several Flash models add a storage fee per million cached tokens per hour. DeepSeek remains the outlier: no write premium, and a V4-Pro cache hit at $0.044, about 3 percent of its peak input price.

Run the numbers on the support-chatbot shape — 20M input and 4M output a month, 75 percent of input served from cache, one-time write premiums excluded — and caching cuts Anthropic's bill from $160 to $103, Astra's from $400 to $265, and DeepSeek's peak bill from $42.24 to $23.10. The ordering does not change. The magnitude does: on agent loops that resend the same system prompt and file context all day, the cache column is routinely worth more than a list-price negotiation would be.

Three Worked Bills at Real Volume

Every bill below applies list-price LLM API pricing with no caching, no batch discounts, and no negotiation — the worst case a brand-new account can hit.

The coding-agent row assumes 300 sessions a day at 60K input and 12K output each. It is also the row where the fine print bites hardest: push any single Astra session past 272K input tokens and that session's inputs bill at 2x and its outputs at 1.5x — a long-context agent can clear $15,000 on the same request count. The workhorse tier tells the opposite story. The identical agent load costs $2,376 on GPT-5.6 Terra, $810 on Gemini 3.8 Flash ($1,620 from January 1), and $291.60 on deepseek-flash at peak, $145.80 off-peak. The spread inside one vendor's menu is now wider than the spread between vendors — which is why model selection, not vendor selection, moves most bills.

When DeepSeek Is the Rational Default — and When It Is Not

DeepSeek's off-peak window deserves a second look from US-based teams, because the peak hours are drawn in UTC: 01:00–04:00 and 06:00–10:00, which maps to 8:00–11:00 PM and 1:00–5:00 AM Eastern. American business hours are entirely off-peak. A US office running batch extraction, evaluation suites, or background summarization during the workday pays the half rate all day, every day — the full picture is on our DeepSeek V4 tool profile.

The honest caveats: 500 concurrent requests on V4-Pro and 2,500 on flash are hard caps that high-fanout agent fleets can hit, the API bills its web-search and file-parsing tools separately — the same unbundling we documented across search APIs — and the standard disclaimer applies with extra force here: a price sheet says nothing about whether the model is good enough for your task. Run both models on your own traffic before migrating anything; a 5x cheaper model that needs 3x the tokens to finish a job is not cheaper. Teams that end up running a mixed fleet — a cheap model for drafts, a frontier one for the hard turn — usually wrap the routing in a gateway rather than hardcoding endpoints; we compared the main options in our LLM gateway roundup.

Frequently Asked Questions

How much does the GPT-6 Astra API cost?

GPT-6 Astra costs $10.00 per million input tokens and $50.00 per million output tokens at standard rates, with cached input at $1.00 per million. Those prices apply up to 272,000 input tokens; beyond that, input bills at 2x ($20.00), cached input at 2x, and output at 1.5x ($75.00). Batch processing halves the short-context rates, and Fast mode — OpenAI's priority lane — runs at double.

Is Claude Opus 5.5 cheaper than GPT-5.6 Sol?

On paper they are identical: $4.00 input and $20.00 output per million tokens. The differences sit underneath — Opus 5.5 reads cached input at $0.20 (5 percent of list) versus Sol's $0.40, and Anthropic publishes no long-context surcharge, while Sol inherits OpenAI's 2x/1.5x multipliers past 272K tokens. Sol's rate is also explicitly promotional through at least November 21, 2026, so the two may not stay matched. At blended list price, both cost $24 per million-in-plus-million-out — the same number Gemini 4 Carbon hits on December 1.

What is the cheapest LLM API in 2026?

For usable quality at scale, deepseek-flash is the cheapest metered option in this comparison: $0.30/$1.20 at peak and $0.15/$0.60 off-peak, with cache hits at $0.006. Anthropic's Haiku 5.5 posts the cheapest single rate at $0.10/$0.50, but only for prompts under 100,000 tokens. OpenAI's Luna ($0.20/$1.20) and Gemini 3.5 Flash-Lite ($0.30/$2.50) bracket the sub-dollar tier. Anything promising frontier quality below these numbers is either an open-weight model on a third-party host or a loss leader — read the throughput and concurrency limits before you migrate.

Why is Gemini 4 pricing increasing on December 1, 2026?

Google launched the Gemini 4 family (Carbon, Argon, Hydrogen) on introductory pricing — $2/$10 for the frontier models — with the increase published on its own pricing page from day one: $4/$20 beginning December 1, 2026, with cache reads doubling to $0.20 the same day. The Flash line follows on January 1, 2027, moving from $0.75/$3.75 to $1.50/$7.50. It is the intro-then-step-up pattern in its purest form, and it means any annual commitment signed in November on Gemini 4 rates should be modeled at the December numbers, not the current ones.

How do DeepSeek's off-peak prices work?

DeepSeek bills deepseek-flash and V4-Pro at two rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays — 35 hours a week. Everything else, including the entire weekend, bills at exactly half: $0.66 input and $1.98 output on V4-Pro, $0.15/$0.60 on flash, with cache hits at $0.022 and $0.003 respectively. Off-peak is not a separate batch queue with worse latency; it is the same synchronous API on the same endpoints, which makes it usable for interactive traffic, not just overnight jobs.

The Bottom Line

Four rate cards, four meters, and no single winner. If you need the highest ceiling regardless of cost, GPT-6 Astra is the tool — just model long-context sessions at the 2x/1.5x multipliers, because that is where its real bill lives. For coding agents and cache-heavy loops, Claude Opus 5.5 at $4/$20 with $0.20 cache reads and stackable 50 percent batch discounts is the most predictable frontier bill in the group. Gemini 4 Carbon is the market's cheapest frontier model for exactly sixteen more days; whatever you build on the $12 blended million, rebuild the budget at $24. And for high-volume, latency-tolerant work — extraction, evaluation, summarization, overnight pipelines — DeepSeek's off-peak meter remains untouchable at $2.64 a blended million. Prices moved in both directions at once this year, which means last quarter's LLM API pricing comparison is already a liability. Check the meters monthly, and route the fleet through a gateway so switching costs stay near zero.