Model comparison
DeepSeek-V4-Pro vs Claude Opus 5.5: where 4.6x comes from
Claude Opus 5.5 costs $4.40 for a cached coding session that costs $0.95 on DeepSeek-V4-Pro. Cache writes and output drive the gap; a V4.1 Pro is planned.
· Prices as of September 28, 2026
DeepSeek-V4-Pro
DeepSeek · Released August 13, 2026
DeepSeek's agent-focused large model, released in general availability in August 2026 with support for OpenAI's Responses API and Codex.
DeepSeek-V4-Pro facts and comparisonsClaude Opus 5.5
Anthropic · Released September 22, 2026
Anthropic's recommended starting model for most work, built for long-running agentic coding. Anthropic says it matches Claude Fable 5.1 on most work at a much lower price.
Claude Opus 5.5 facts and comparisons
The short answer
DeepSeek-V4-Pro costs $0.95 for the example agentic coding session against $4.40 on Claude Opus 5.5, a 4.6x gap in which cache writes account for $2.07 and output, at $20 against $3.96 per million, for most of the rest. Claude Opus 5.5 is the default model in Claude Code and Anthropic's recommended starting model for most work. DeepSeek-V4-Pro suits cost-driven agent work through OpenRouter, OpenCode, or self-hosting, though DeepSeek says a V4.1 Pro model is on the way.
Choose DeepSeek-V4-Pro if
- Output dominates your spend: DeepSeek-V4-Pro charges $3.96 per million output tokens at peak against $20, so the output-heavy generation costs $0.36 instead of $1.72.
- You want open weights under the MIT license, with DeepSeek-V4-Pro-0813 as the current snapshot.
- You want explicit reasoning effort levels, which DeepSeek lists as low, high, and max for V4-Pro.
- You need up to 384K output tokens in one response, 3x the 128K Opus 5.5 allows.
Choose Claude Opus 5.5 if
- Claude Code is your daily tool and you want its default model, Opus 5.5 on Pro, Max, Team, and Enterprise plans and with an Anthropic API key.
- Your jobs look like the ones Anthropic highlights for Opus 5.5: codebase-wide migrations, audits, and other long, sprawling work.
- You want Opus 5.5's fast mode, a Claude API research preview billed at $8 input and $40 output per million.
- You also use Cursor or GitHub Copilot, which offer Opus 5.5 and do not list DeepSeek-V4-Pro.
Side by side
Specs and prices
| Fact | DeepSeek-V4-Pro | Claude Opus 5.5 |
|---|---|---|
| Maker | DeepSeek | Anthropic |
| API model id | deepseek-v4-pro | claude-opus-5-5 |
| Released | August 13, 2026 | September 22, 2026 |
| Status | Current | Current |
| Context window | 1M tokens | 1M tokens |
| Max output | 384K tokens | 128K tokens |
| Open weights | Yes | No |
| Input, per 1M tokens | $1.32 | $4 |
| Cache hit, per 1M | $0.044 | $0.20 |
| Cache write, per 1M | $1.32 (same as input) | $5 (5-minute), $8 (1-hour) |
| Output, per 1M tokens | $3.96 | $20 |
| Runs in | OpenCode and OpenRouter | Claude Code, Cursor, OpenCode, OpenRouter, and GitHub Copilot |
Standard API rates in US dollars, as published by each maker (DeepSeek-V4-Pro: September 28, 2026; Claude Opus 5.5: September 26, 2026). Batch and priority tiers, taxes, and subscription plans are not included. DeepSeek-V4-Pro: Prices are DeepSeek's peak rates. Off-peak hours cost 50% less: peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Claude Opus 5.5: The full 1M context window is billed at standard rates. Fast mode, a research preview on the Claude API, costs $8 input and $40 output per million tokens.
Cost
What the same work costs
The same token counts, priced at each model’s published rates. Cached tokens are billed at each maker’s cache prices, so the session shows what caching is worth on each model.
Real sessions differ: the two models count the same code as different numbers of tokens, and reasoning settings change how much each one writes. Your own history is the real test.
| Workload | DeepSeek-V4-Pro | Claude Opus 5.5 |
|---|---|---|
| Agentic coding session, 100K input, 400K written to cache, 2M read from cache, 50K output | $0.95 | $4.40 |
| Large one-off review, 150K input with no cache hits, 10K output | $0.24 | $0.80 |
| Output-heavy generation, 30K input, 80K output | $0.36 | $1.72 |
| A month of sessions, 110 sessions: 5 a day, 22 working days | $104.06 | $484.00 |
| Where the session’s cost goes | ||
| Cache writes | $0.53 | $2.60 |
| Cache reads | $0.09 | $0.40 |
| Uncached input | $0.13 | $0.40 |
| Output | $0.20 | $1.00 |
- caching saves on the session with DeepSeek-V4-Pro (73%)
- $2.55
- caching saves on the session with Claude Opus 5.5 (60%)
- $6.60
Without the cache, output rates set the gap
Per million tokens, Claude Opus 5.5 costs 3x as much as DeepSeek-V4-Pro for input, $4 against $1.32, and 5.1x as much for output, $20 against $3.96. The DeepSeek rates are its peak rates. The large one-off review, mostly input, shows a 3.3x gap, $0.80 against $0.24, while the output-heavy generation shows 4.8x, $1.72 against $0.36.
That output ratio is the one to watch, because reasoning settings change how many output tokens a task produces. Opus 5.5 runs adaptive thinking at medium effort by default, and changing effort keeps its prompt cache. DeepSeek lists low, high, and max effort for V4-Pro. Each maker defines its own effort levels, and they do not map one to one, so the same task can write different amounts of output on each.
Cache pricing on Opus 5.5 and DeepSeek-V4-Pro
Anthropic prices a cache hit on Opus 5.5 at 0.05x input, 5% of its input price, where most Claude models use 0.1x. That makes an Opus 5.5 hit $0.20 per million. DeepSeek goes lower still: $0.044 on V4-Pro, 3.3% of input, and its disk cache is on by default for every account with no code changes. On cache reads the gap is 4.5x, narrower than on output.
Writes separate them more. DeepSeek lists no fee for writing the cache, so the 400K tokens the example session writes cost ordinary input, $0.53. Anthropic charges 1.25x input for a 5-minute write and 2x for a 1-hour write, $5 and $8 per million on Opus 5.5, so the same writes cost $2.60, 59% of its session. At $2.07, writes are the largest single piece of the $3.45 gap. Which of those two rates Claude Code pays depends on how you sign in: its main conversation writes to the 1-hour cache on a subscription and to the 5-minute cache with an API key.
The totals come to $0.95 on DeepSeek-V4-Pro and $4.40 on Opus 5.5 per session, or $104.06 against $484.00 over 110 sessions a month. DeepSeek bills 50% less outside its peak hours, 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so off-peak V4-Pro work costs less than the table shows. The cache math on the homepage walks through the same session.
What Anthropic and DeepSeek say, and what comes next
Anthropic calls Opus 5.5 its recommended starting model for most work and says "It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." It names agentic coding, computer use, and knowledge work as strengths, and claims output more than 30% faster than Claude Opus 5.
DeepSeek says "The GA version of DeepSeek V4 Pro greatly enhances agent capabilities, with particularly significant performance improvements in production environments." It also says V4-Pro natively supports OpenAI's Responses API and is adapted for Codex with one-click setup. Two caveats come from DeepSeek itself: service continues with billing unchanged until a V4.1 Pro model arrives, and DeepSeek says its newer, cheaper DeepSeek-V4.1-Flash is ahead of V4-Pro on performance, cost, speed, and total runtime in tests by several parties.
Where they run differs too. Claude Code is Anthropic's own agent and runs Claude models, and Opus 5.5 is also in Cursor, OpenRouter, OpenCode, and GitHub Copilot. DeepSeek-V4-Pro is in OpenRouter and OpenCode. On OpenRouter a request for DeepSeek's open weights goes to one of several providers, whose prices can differ from the DeepSeek API price used here, and self-hosting the weights is possible at a cost this page does not estimate.
EveryToken prices Opus 5.5 at Anthropic's rates from Claude Code, Cursor, or OpenCode history, and DeepSeek-V4-Pro when you run it through OpenRouter, so both show up on your own sessions.
Prompt caching
How each maker bills cached tokens
DeepSeek
DeepSeek's disk cache is on by default for every account, with no code changes. DeepSeek lists no fee for writing the cache.
A cache hit costs $0.006 per million tokens on DeepSeek-V4.1-Flash and $0.044 on DeepSeek-V4-Pro at peak rates, and off-peak hours cost 50% less.
Source: DeepSeek API: Context caching
Anthropic
Claude caches a prompt prefix up to a breakpoint. One top-level cache_control field places the breakpoint automatically and moves it as the conversation grows, or you can mark up to 4 blocks yourself. Claude Code manages caching for you.
A 5-minute cache write costs 1.25x the input price and a 1-hour write costs 2x. A cache hit costs 0.1x input on most models, 0.05x on Claude Opus 5.5, and 0.025x on Claude Fable 5.1 and Claude Mythos 5.1. Every hit restarts the cache lifetime at no charge.
Anthropic's rule of thumb: a 5-minute write pays for itself after one cache read, and a 1-hour write after two.
In Claude Code, the main conversation uses the 1-hour cache on a Claude subscription and the 5-minute cache with an API key. Each model has its own cache, so switching models starts over.
Source: Anthropic: Prompt caching
Your own numbers
See what DeepSeek-V4-Pro and Claude Opus 5.5 really cost you.
everyaitoken reads your OpenRouter, Claude Code, Cursor, and OpenCode history on your Mac and prices every request at API rates, with what caching saved or cost. $9 once.
FAQ
Questions
How much cheaper is DeepSeek-V4-Pro than Claude Opus 5.5?
At DeepSeek's peak rates, 3x on input and 5.1x on output. The example cached session costs $0.95 against $4.40, a 4.6x gap, and 110 sessions come to $104.06 against $484.00.
Is DeepSeek-V4-Pro being replaced?
DeepSeek says V4-Pro service continues with billing unchanged until a V4.1 Pro model arrives. It has already released DeepSeek-V4.1-Flash, which it says outperforms V4-Pro in tests by several parties.
Which model does Claude Code use by default?
Opus 5.5 is what Claude Code starts with on Pro, Max, Team, and Enterprise plans, and with an Anthropic API key. Claude Code is built around Anthropic's models and can reach other providers' compatible endpoints only with custom configuration, which this post doesn't cover. OpenCode and OpenRouter both list DeepSeek-V4-Pro.
Can I self-host DeepSeek-V4-Pro?
Yes, its weights are open under the MIT license. The costs on this page are DeepSeek's own API prices and do not cover the hardware or hosting a self-hosted deployment needs.
Sources
- DeepSeek API: Models and pricing
- DeepSeek API: Change log
- DeepSeek: V4 Pro release
- OpenRouter: DeepSeek-V4-Pro-0813
- OpenCode docs: Zen
- Anthropic: Pricing
- Anthropic docs: Claude Opus 5.5
- Anthropic: Claude Opus 5.5
- Anthropic docs: Fast mode
- Claude Code docs: Model configuration
- Cursor docs: Claude Opus 5.5
- OpenRouter: Claude Opus 5.5
- OpenCode docs: Zen
- GitHub Docs: Supported AI models in Copilot
- DeepSeek API: Context caching
- Anthropic: Prompt caching