Claude Mythos Pricing: What the Capybara Tier Will Cost
Anthropic has not published official pricing for Claude Mythos, but the company has confirmed the model is “very expensive for us to serve, and will be very expensive for our customers to use.” Based on Anthropic’s current pricing structure, leaked internal documents, and the historical pattern of AI model pricing, Capybara-tier pricing is expected to land between 2x and 5x current compared to Opus rates — meaning $10-25 per million input tokens and $50-125 per million output tokens.
Current Claude API Pricing Baseline
All Model Tiers Compared
Understanding Claude Mythos pricing starts with knowing what Anthropic charges today. The company operates a four-tier pricing structure that directly correlates capability with cost. Every tier uses the same Messages API, so the only difference between a $1/MTok Haiku call and a $5/MTok Opus call is the model parameter string.
| Model | Input | Output | Batch Input | Batch Output | Context |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | $1/MTok | $5/MTok | $0.50/MTok | $2.50/MTok | 200K |
| Claude Sonnet 4.6 | $3/MTok | $15/MTok | $1.50/MTok | $7.50/MTok | 200K-1M |
| Claude Opus 4.6 | $5/MTok | $25/MTok | $2.50/MTok | $12.50/MTok | 200K-1M |
| Claude Mythos (est.) | $10-25/MTok | $50-125/MTok | $5-12.50/MTok | $25-62.50/MTok | 1M+ |
The output-to-input pricing ratio has been consistent at 5:1 across all Claude models. Haiku charges $1/$5, Sonnet charges $3/$15, and Opus charges $5/$25. If this ratio holds for Capybara, an input price of $15/MTok would mean $75/MTok output — matching the pricing of the now-deprecated Claude Opus 4.1.
Opus 4.6 as the Pricing Benchmark
Claude Opus 4.6 is the right baseline for estimating Mythos pricing. At $5 per million input tokens and $25 per million output tokens, Opus 4.6 already sits at the premium end of the AI API market. For context, Opus 4.6 represents a significant price drop from earlier Opus generations — Claude Opus 4.1 and Opus 4 both cost $15 input and $75 output, three times more than the current flagship.
This price reduction history matters. Anthropic has shown willingness to launch models at premium prices and reduce them over time as serving efficiency improves. Mythos will likely follow the same trajectory: launch expensive, then drop as Anthropic optimizes inference costs.
Claude Mythos Pricing Estimates
Conservative Estimate: 2-3x Opus
The conservative pricing scenario puts Capybara at $10-15 per million input tokens and $50-75 per million output tokens. This is grounded in several data points. The jump from Sonnet to Opus is roughly 1.7x for input tokens. A similar multiplier from Opus to Capybara would put input pricing around $8.50-10/MTok.
At $15 input and $75 output, Mythos would match the exact pricing of the previous-generation Claude Opus 4.1 — a model that Anthropic already proved the market would pay for. This price point makes strategic sense: it’s expensive enough to reflect premium capability, but not so expensive that it kills adoption before the October 2026 IPO.
The batch API at 50% discount would bring these costs to $5-7.50 input and $25-37.50 output — essentially matching current Opus standard pricing for workloads that can tolerate async processing.
Aggressive Estimate: 4-5x Opus
The aggressive pricing scenario puts Capybara at $20-25 per million input tokens and $100-125 per million output tokens. This estimate comes from Anthropic’s own characterization of the model as “very expensive for us to serve.” If Mythos requires significantly more compute per token than Opus — which the “larger” description suggests — Anthropic may need higher pricing to maintain margins.
At $25/$125 per million tokens, a single Mythos API call processing 100K input tokens and generating 4K output tokens would cost approximately $3.00 — compared to $0.60 for the same call on Opus 4.6. For applications making hundreds of API calls daily, this five-fold increase would fundamentally change the economics.
What Anthropic Has Said About Cost
Anthropic’s leaked materials contain two direct references to cost. First, the model is described as “very expensive for us to serve, and will be very expensive for our customers to use.” Second, Anthropic is “working to make it much more efficient before any general release.” These statements together suggest launch pricing will be at the high end of estimates, with planned reductions as inference optimization progresses.
The efficiency work is not unusual. Every large language model becomes cheaper to serve over time through techniques like quantization, speculative decoding, and hardware optimization. Opus 4.6 at $5/MTok is three times cheaper than Opus 4.1 at $15/MTok, largely due to these optimizations applied over successive generations.
Full Pricing Comparison Table
Input and Output Pricing Across All Tiers
Here is the complete pricing picture for all Claude models, including estimated Capybara pricing at both conservative and aggressive estimates:
| Model | Standard Input | Standard Output | Cache Hits | 5m Cache Write | 1h Cache Write |
|---|---|---|---|---|---|
| Haiku 4.5 | $1.00 | $5.00 | $0.10 | $1.25 | $2.00 |
| Sonnet 4.6 | $3.00 | $15.00 | $0.30 | $3.75 | $6.00 |
| Opus 4.6 | $5.00 | $25.00 | $0.50 | $6.25 | $10.00 |
| Mythos (low est.) | $10.00 | $50.00 | $1.00 | $12.50 | $20.00 |
| Mythos (high est.) | $25.00 | $125.00 | $2.50 | $31.25 | $50.00 |
Cache hit pricing at 10% of base input cost is the single most important cost lever for Capybara. At the conservative estimate, a cache hit costs just $1/MTok — the same as Haiku 4.5 standard input. At the aggressive estimate, cache hits at $2.50/MTok still undercut Sonnet 4.6 standard pricing.
Batch API Pricing
The Batch API provides a flat 50% discount on all token pricing. For Capybara, this means:
| Estimate | Batch Input | Batch Output | Equivalent To |
|---|---|---|---|
| Conservative ($10/$50) | $5.00/MTok | $25.00/MTok | Current Opus standard |
| Mid-range ($15/$75) | $7.50/MTok | $37.50/MTok | Former Opus 4.1 batch |
| Aggressive ($25/$125) | $12.50/MTok | $62.50/MTok | Former Opus 4.1 standard |
Batch processing is the great equalizer. If Mythos launches at $15/$75 (the most commonly cited estimate), batch pricing of $7.50/$37.50 puts it between current Opus and former Opus 4.1 pricing — a range the market has already validated.
How to Reduce Claude Mythos Costs
Prompt Caching Strategy
Prompt caching is the highest-impact cost optimization for Capybara-tier pricing. Cache hits cost just 10% of standard input pricing — turning a $25/MTok input into $2.50/MTok on repeat requests. For applications with stable system prompts, document context, or conversation history, caching can reduce total costs by 60-80%.
Two caching options exist: 5-minute cache writes at 1.25x base input price, and 1-hour cache writes at 2x base input price. The 5-minute cache pays for itself after just one cache hit. The 1-hour cache requires two hits to break even but is better for applications with intermittent usage patterns.
For a system prompt of 10,000 tokens accessed 50 times per hour, the math works out clearly. Without caching: 50 × 10,000 × $25/MTok = $12.50/hour. With 1-hour caching: one write at $50/MTok (10,000 × $50/MTok = $0.50) plus 49 hits at $2.50/MTok (49 × 10,000 × $2.50/MTok = $1.225) = $1.725/hour — an 86% reduction.
Batch API for Non-Realtime Workloads
Any workload that doesn’t require synchronous responses should use the Batch API. The 50% discount applies to both input and output tokens, and batch requests can be queued and processed within a 24-hour window. Common use cases include document analysis, code review, content generation, and data extraction.
At estimated Capybara pricing of $15/$75, the batch discount brings costs to $7.50/$37.50 — still premium, but manageable for high-value tasks. Combine batch processing with prompt caching for maximum savings.
Model Routing: Use Capybara Only When Needed
The smartest cost strategy is not to use Mythos for everything. Build a model router that sends simple queries to Haiku ($1/$5), complex queries to Sonnet ($3/$15), and only the most demanding tasks to Capybara. This approach can reduce total API spend by 70-90% compared to routing all traffic through the most expensive model.
A practical routing strategy for a typical application:
- 80% of requests → Haiku 4.5 (classification, extraction, simple Q&A): ~$1-5/MTok
- 15% of requests → Sonnet 4.6 (reasoning, analysis, content): ~$3-15/MTok
- 5% of requests → Capybara/Mythos (complex reasoning, cybersecurity, multi-step agents): ~$10-75/MTok
The blended cost with this distribution stays close to Sonnet pricing even while accessing Capybara capabilities for the hardest tasks.
Fast Mode vs Standard Mode
Claude Opus 4.6 already offers a Fast Mode at 6x standard pricing ($30/$150 per MTok) for significantly faster output. If Capybara offers a similar fast mode, expect pricing of $60-150 input and $300-750 output per million tokens — extreme numbers suited only for latency-critical, high-value applications.
For most use cases, standard mode with prompt caching will deliver better cost efficiency than fast mode. Reserve fast mode for real-time applications where response latency directly impacts revenue or user experience.
Claude Mythos Enterprise Pricing
Volume Discounts
Anthropic offers negotiated volume discounts for high-usage enterprise customers. These discounts are arranged case-by-case through the sales team and can significantly reduce effective per-token costs. For organizations expecting to spend $10,000+ per month on Capybara API usage, enterprise pricing discussions should begin before the model launches.
Contact sales@anthropic.com or the Claude Console to initiate enterprise pricing conversations. Having usage projections ready — estimated monthly token volume, primary use cases, expected traffic patterns — will accelerate the negotiation.
Custom Rate Limits
Enterprise accounts can negotiate custom rate limits beyond the standard Tier 4 caps. For Capybara specifically, custom rate limits may be the only way to access meaningful throughput at launch, since initial rate limits for a new tier are typically conservative.
Third-Party Platform Pricing
AWS Bedrock, Google Vertex AI, and Microsoft Foundry each set their own pricing for Claude models. Regional endpoints include a 10% premium over global endpoints on both Bedrock and Vertex AI. At Capybara pricing levels, this 10% premium becomes significant — an extra $1-2.50 per million input tokens on top of already-premium base pricing.
Third-party platforms may also offer committed-use discounts, reserved capacity pricing, or bundled pricing with other cloud services. Evaluate all three platforms when Capybara becomes available, as pricing differences at this tier translate to thousands of dollars monthly for production workloads.
Cost Calculator: Mythos vs Opus vs Sonnet
Light Usage: 10M Tokens/Month
For a startup or small team making roughly 500 API calls daily with average 1K input and 500 output tokens per call:
| Model | Monthly Input Cost | Monthly Output Cost | Total |
|---|---|---|---|
| Sonnet 4.6 | $22.50 | $56.25 | $78.75 |
| Opus 4.6 | $37.50 | $93.75 | $131.25 |
| Mythos (conservative) | $75.00 | $187.50 | $262.50 |
| Mythos (aggressive) | $187.50 | $468.75 | $656.25 |
Medium Usage: 100M Tokens/Month
For a growing SaaS product processing documents, running agents, and handling customer queries:
| Model | Monthly Input Cost | Monthly Output Cost | Total |
|---|---|---|---|
| Sonnet 4.6 | $225 | $562.50 | $787.50 |
| Opus 4.6 | $375 | $937.50 | $1,312.50 |
| Mythos (conservative) | $750 | $1,875 | $2,625 |
| Mythos (aggressive) | $1,875 | $4,687.50 | $6,562.50 |
Heavy Usage: 1B Tokens/Month
For enterprise-scale applications, cybersecurity platforms, or high-volume agent systems:
| Model | Monthly Input Cost | Monthly Output Cost | Total |
|---|---|---|---|
| Sonnet 4.6 | $2,250 | $5,625 | $7,875 |
| Opus 4.6 | $3,750 | $9,375 | $13,125 |
| Mythos (conservative) | $7,500 | $18,750 | $26,250 |
| Mythos (aggressive) | $18,750 | $46,875 | $65,625 |
At heavy usage, the difference between conservative and aggressive Mythos estimates is nearly $40,000/month. This underscores why organizations should use model routing — sending only the hardest 5-10% of queries to Capybara while handling the rest with cheaper tiers.
Questions About Claude Mythos Pricing
How much will Claude Mythos cost per token?
Official pricing has not been announced. Conservative estimates place it at $10-15 per million input tokens and $50-75 per million output tokens. Aggressive estimates range up to $25 input and $125 output per million tokens. Anthropic confirmed the model will be “very expensive.”
Is Claude Mythos more expensive than GPT-4?
Yes. Even at conservative estimates, Mythos at $10-15/MTok input significantly exceeds most competing models. However, Anthropic positions Capybara as a breakthrough tier for tasks where no other model delivers adequate results — particularly cybersecurity and complex multi-step reasoning.
Will the batch API discount apply to Claude Mythos?
Almost certainly. The 50% batch discount applies to all current Claude models, and Anthropic has not indicated any change to this policy. Batch pricing for Mythos would be $5-12.50 input and $25-62.50 output per million tokens depending on the base price.
How can I reduce Claude Mythos API costs?
Three strategies: prompt caching (reduces repeat input costs by 90%), batch processing (50% discount on all tokens), and model routing (send only complex tasks to Capybara, use Haiku or Sonnet for everything else). Combined, these strategies can reduce effective Mythos costs by 70-90%.
When will Anthropic announce official Mythos pricing?
No date confirmed. Pricing will be announced alongside public API availability. Anthropic is “working to make it much more efficient before any general release,” suggesting pricing may be lower at launch than current estimates if optimization succeeds.
Does prompt caching work with Claude Mythos?
Pricing details for Capybara caching are unconfirmed, but all current Claude models support prompt caching with cache hits at 10% of standard input pricing. If this multiplier applies to Capybara, cache hits would cost $1-2.50/MTok — comparable to Haiku or Sonnet standard pricing.
How does Claude Mythos pricing compare to Opus 4.6?
Mythos is expected to cost 2-5x more than Opus 4.6. Opus 4.6 charges $5 input and $25 output per million tokens. Capybara estimates range from $10-25 input and $50-125 output. The actual multiplier depends on model size and Anthropic’s inference optimization progress.
Will there be a free tier for Claude Mythos?
Unlikely at launch. Anthropic provides small free credits for new API accounts, but these typically cover only lower-tier models. Given the high serving cost of Capybara, free access would be economically unsustainable. Enterprise trials through the sales team may be available.