Claude Opus 5 pricing: real cost per million tokens
What Anthropic's Claude Opus 5 really costs across 5 plans from 2 providers: monthly fees, usable allowances and real per-token prices, against the official API list price and independent leaderboards.
Data updated · How real prices are computed
As of 2026-10-08, the cheapest way to use Claude Opus 5 is Claude Pro at a real cost of about $0.00772 per million tokens ($20/month for roughly 2.589B tokens a month). We track 5 plans and API channels that serve Claude Opus 5. Official API list price per million tokens: $0.5 cached input, $5 input, $25 output — about $0.735 per million tokens on a standard workload mix.
- Lowest real price
- $0.00772/MTokClaude Pro
- API list price
- $0.735/MTokstandard workload mix
- Most monthly tokens
- 15.7BClaude Max 20x (9/14+)
- Plans & APIs
- 54 subscriptions · 1 metered API
Plot Claude Opus 5 on the price–capability chart →
Claude Opus 5 plans and real prices
Real price = monthly fee ÷ tokens you can actually use in a month. “vs API” shows how many times cheaper than the official API list price (standard workload mix) the plan is. “Evidence” opens the full calculation and sources for that row.
| Plan | Provider | Billing | Monthly fee | Monthly tokens | Real $/MTok | vs API | Confidence | Data date | |
|---|---|---|---|---|---|---|---|---|---|
| Claude Pro | Anthropic | Subscription | $20/mo | 2.589B | $0.00772 | 95.1× | High | Evidence ↓ | |
| Claude Max 5x (9/14+) | Anthropic | Subscription | $100/mo | 7.85B | $0.01274 | 57.7× | Medium | Evidence ↓ | |
| Claude Max 20x (9/14+) | Anthropic | Subscription | $200/mo | 15.7B | $0.01274 | 57.7× | Medium | Evidence ↓ | |
| Droid Max | Factory | Subscription | $200/mo | 2.021B | $0.09897 | 7.4× | Medium | Evidence ↓ | |
| Claude Opus 5 API | Anthropic | Metered API | — | — | $0.7663 | — | High | Evidence ↓ |
Behind each number
For every plan: the data date, the exact calculation, the quota basis, how leaderboard scores were mapped, and the original adoption evidence with its sources.
Claude ProAnthropic$0.00772499/MTok
- Plan
- Claude Pro · Anthropic
- Data date
- 2026-09-14 – 2026-09-22 · sampled
- Real price / MTok
- $0.00772499
- Monthly tokens
- 2,589,000,000 (2.589B)
- Quota confidence
- High
- Calculation
- $20 / month ÷ 2,589,000,000 tokens × 1,000,000 ≈ $0.00772499 / MTok
- Quota basis
Anthropic workload: 97% cache reads / 2.5% cache writes / 0.5% output
Official API list (cached / in / out, per MTok): $0.5 / $5 / $25 · standard-load blended ≈ $0.735/MTok
- Benchmark mapping
- Artificial Analysis Intelligence Index v4.3: Medium (5 configurations)
Terminal-Bench 4.0: Low (5 configurations)
Terminal-Bench 4.0 (AA): Low (5 configurations)
Code Arena | WebDev Overall: Low (2 configurations)
Agent Arena · Overall (Ranking / Pareto): Low (2 configurations)
Artificial Analysis Coding Agent Index v1.5: Low (1 configuration)
DeepSWE v1.1: Low (5 configurations)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
issue #53 (NTRYourWaifu): two Pro accounts polling /api/oauth/usage seven_day % on 9/14~9/22, deduplicated by message.id — Opus 5 total 1,022.02M raw = +148pp of the week; community-issues-round1-2026-09-29.json; the old source X @shownotover Pro /usage panel 32.87M = 7% of the weekly pool (claude-adoption-round8-2026-09-20.json) is no longer weighted; workload-conversion-round1-2026-09-30.json
2.762B→2.589B (2026-09-30 decision: uniform workload-normalization for measured samples with token breakdown, Anthropic tier): segment worth $733.93 (read 1,004.96M×$0.5 + 1h write 12.99M×$10 + in 0.01M×$5 + out 4.06M×$25) = $4.96 per 1% → $495.9/week ×4 weeks ÷ Anthropic-tier blended price $0.76625/MTok; the raw-basis 27.62 is kept for comparison. The issue #53 adoption decision — 1.878B→2.762B, 2026-09-30 decision switching to issue #53: two Pro accounts, all after the permanent 9/14 +25%: 1,022.02M raw (read 1,004.96M / 1h write 12.99M / in 0.01M / out 4.06M) = +148 weekly percentage points → 6.906M/pp ×100 ×4 weeks; the two full weeks give 2.460B and 2.896B respectively; worth per 1% at list prices is $4.96. The old panel sample gives $3.67 per 1% (7% integer reading, interval 1.753B~2.023B) — 35%/47% higher, close to the +25% adjustment, presumably measured before 9/14, so it is not merged (the merged 2.722B is not adopted); the Anthropic-tier worth basis is 2.589B (ruled adopted from 2026-09-30, see first line). Tension: the ratio to the Max 20x adopted value 157 changes from 8.35× to 5.68×, inconsistent with Opus 5.5's 301.7/30.54≈9.9× — pointing to the Max 20x basis or the 5.5/5 pool weights (#52/#53 derive ≈0.843), pending ruling. Original old 18.78 decision text — Opus 4.8's historical 1.588B → Opus 5 panel back-derivation 1.878B: 32,868,513 total (opus5 32.85M + haiku 23k) ÷7%×4 weeks; weekly pool 470M, 5h pool 56.7M — the weekly pool is the binding constraint; weekly-% rounding interval about 1.75B~2.02B; single session n=1, initially medium; round5's Opus5 candidate ≈190M (assuming weekly message counts) was voided by the direct panel measurement; the list-price closure check $25.68 agrees; 2026-09-21 decision: Opus 4.8's old measurement gets no weight — round8 had used the baseline-period reading ×1.25 = 1.985B as corroboration, but round6 records that value as containing a +50% promo boost; the faithful mapping ÷1.5×1.25 = 1.323B contradicts the panel, so only the panel source stands; round9's second post by the same author, 40M = 13% of the week, is recorded as tension: new-week reading 1.23B / continuing-week reading 2.67B / lower-bound reading non-binding — the baseline is unknowable, not weighted; 2026-09-21 decision raising to high: panel-grade form + price closure check + post-cut basis + self-consistent 8.35× weekly-pool ratio vs Max 20x; #3's new-week reading 1.23B would unbalance the Max pool ratio (a $20 tier getting a $100 tier's relative pool share is abnormal), corroborating the 18.78 side. Note: '8.28 windows/week isomorphic to Max' was once cited as corroboration; after round9 checkpoints back-derived the Max tier's 5h pool at ~710M (not the 474M full-window reading), that corroboration was withdrawn — Max's full windows per week are ~5.5, not 8.3
claude-adoption-round8-2026-09-20.jsoncommunity-issues-round1-2026-09-29.jsonworkload-conversion-round1-2026-09-30.jsonBrowse the public evidence archive
Claude Max 5x (9/14+)Anthropic$0.0127389/MTok
- Data date
- 2026-08 · derived; uses the data date of Claude Max 20x (9/14+) · claude-opus-5
- Real price / MTok
- $0.0127389
- Monthly tokens
- 7,850,000,000 (7.85B)
- Quota confidence
- Medium
- Calculation
- $100 / month ÷ 7,850,000,000 tokens × 1,000,000 ≈ $0.0127389 / MTok
- Quota basis
Not workload-normalized: sample lacks a token breakdown (or is an official absolute token table); raw tokens used as-is
Official API list (cached / in / out, per MTok): $0.5 / $5 / $25 · standard-load blended ≈ $0.735/MTok
- Benchmark mapping
- Artificial Analysis Intelligence Index v4.3: Medium (5 configurations)
Terminal-Bench 4.0: Low (5 configurations)
Terminal-Bench 4.0 (AA): Low (5 configurations)
Code Arena | WebDev Overall: Low (2 configurations)
Agent Arena · Overall (Ranking / Pareto): Low (2 configurations)
Artificial Analysis Coding Agent Index v1.5: Low (1 configuration)
DeepSWE v1.1: Low (5 configurations)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
Confirmed the 20x weekly pool is exactly 2× the 5x pool; claude-adoption-round6-2026-09-06.json
Old 3.56B→7.85B, permanent basis from 9/14: 157÷2; low→medium, inferred from the confirmed weekly-pool relation — not an independent measurement; neither the 3.6B message-count candidate nor 7B / the old 2.25× multiplier adopted; the 4× relation of the 5h window is not applied to the weekly pool
claude-adoption-round6-2026-09-06.jsonBrowse the public evidence archive
Claude Max 20x (9/14+)Anthropic$0.0127389/MTok
- Data date
- 2026-08 · sampled
- Real price / MTok
- $0.0127389
- Monthly tokens
- 15,700,000,000 (15.7B)
- Quota confidence
- Medium
- Calculation
- $200 / month ÷ 15,700,000,000 tokens × 1,000,000 ≈ $0.0127389 / MTok
- Quota basis
Not workload-normalized: sample lacks a token breakdown (or is an official absolute token table); raw tokens used as-is
Official API list (cached / in / out, per MTok): $0.5 / $5 / $25 · standard-load blended ≈ $0.735/MTok
- Benchmark mapping
- Artificial Analysis Intelligence Index v4.3: Medium (5 configurations)
Terminal-Bench 4.0: Low (5 configurations)
Terminal-Bench 4.0 (AA): Low (5 configurations)
Code Arena | WebDev Overall: Low (2 configurations)
Agent Arena · Overall (Ranking / Pareto): Low (2 configurations)
Artificial Analysis Coding Agent Index v1.5: Low (1 configuration)
DeepSWE v1.1: Low (5 configurations)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
Zenn skipbit measurement + maintainer's permanent-basis figure; claude-adoption-round6-2026-09-06.json
Old 8B→15.7B, permanent basis from 9/14: 4.72B/week ×4 weeks ÷1.5 ×1.25, then rounded; reference interval 11B~20B. Mixed-model and not-fully-same-window samples — not a hard cap from pure Opus 5 measurement; neither the promo-period 189 nor the bare-baseline 126 adopted; round9 alldonesites pure-Opus5 full-window median 474M (cluster 415M~474M across three sources) is a window-capacity reading — the 5h-pool basis remains disputed: chudi checkpoint back-derivation ~710M vs the official 20× Pro implying 1.134B, and Max's full windows per week are ~5.5 rather than the earlier-noted 8.3, so window structure is not used as adoption evidence; after the round9 timeline correction the Reddit audit turns positive: "this week" 2.1B cache-read on one account = 52% of the week is a post-9/17-reset post-cut reading → weekly pool ≈4.0B → 16.1B/month, within 3% of the adopted value — the best weekly-pool corroboration for a pure Opus workload; "last week" 23.7B raw = 230% fits the promo-period large-pool basis and is no longer contradictory; the only remaining low-end tension is its 5-window assumption of 9.5B; not workload-normalized: the Zenn mixed-model sample has no token breakdown, kept raw
claude-adoption-round6-2026-09-06.jsonBrowse the public evidence archive
Droid MaxFactory$0.0989707/MTok
- Data date
- 2026-09-29 · derived; uses the data date of Droid Max · claude-opus-5.5
- Real price / MTok
- $0.0989707
- Monthly tokens
- 2,020,800,000 (2.021B)
- Quota confidence
- Medium
- Calculation
- $200 / month ÷ 2,020,800,000 tokens × 1,000,000 ≈ $0.0989707 / MTok
- Quota basis
Anthropic workload: 97% cache reads / 2.5% cache writes / 0.5% output
Official API list (cached / in / out, per MTok): $0.5 / $5 / $25 · standard-load blended ≈ $0.735/MTok
- Benchmark mapping
- Artificial Analysis Intelligence Index v4.3: Medium (5 configurations)
Terminal-Bench 4.0: Low (5 configurations)
Terminal-Bench 4.0 (AA): Low (5 configurations)
Code Arena | WebDev Overall: Low (2 configurations)
Agent Arena · Overall (Ranking / Pareto): Low (2 configurations)
Artificial Analysis Coding Agent Index v1.5: Low (1 configuration)
DeepSWE v1.1: Low (5 configurations)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
Derived from same-plan claude-opus-5.5 2.526B × 0.8
Factory official multipliers Opus 5.5 1.6× / claude-opus-5 2×: the same Standard Usage pool scaled by multiplier — not a measurement of this model; inherits the Opus 5.5 row's Anthropic-tier workload, not recomputed for each model's own price gap; docs.factory.ai/docs/models; droid-model-multipliers-2026-09-29.json
droid-model-multipliers-2026-09-29.jsonBrowse the public evidence archive
Claude Opus 5 APIAnthropic$0.76625/MTok
- Plan
- Claude Opus 5 API · Anthropic
- Data date
- 2026-09-05 · official source date
- Real price / MTok
- $0.76625
- Quota confidence
- High
- Calculation
- Metered API · public token prices weighted by the Anthropic workload (input share at the 5-minute cache-write rate).
- Benchmark mapping
- Artificial Analysis Intelligence Index v4.3: Medium (5 configurations)
Terminal-Bench 4.0: Low (5 configurations)
Terminal-Bench 4.0 (AA): Low (5 configurations)
Code Arena | WebDev Overall: Low (2 configurations)
Agent Arena · Overall (Ranking / Pareto): Low (2 configurations)
Artificial Analysis Coding Agent Index v1.5: Low (1 configuration)
DeepSWE v1.1: Low (5 configurations)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
platform.claude.com/docs/en/about-claude/pricing
List price cached 0.5/in 5.0/out 25.0, cache write (5m) $6.25 × Anthropic unified workload 97.0%/2.50%/0.50% (the regular-input share is billed at the 5-minute cache-write price)
Claude Opus 5 benchmark scores
Highest archived configuration per leaderboard; ranks are among all models each board archives in this dataset. Scores are never mixed across boards.
| Leaderboard | Metric | Score | Rank | Configuration | Effort | Harness | Configs | Snapshot |
|---|---|---|---|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3 | Intelligence Index | 50.78 | #6 / 60 | Claude Opus 5 (max) | Max | — | 5 | 2026-10-01 |
| Terminal-Bench 4.0 | Resolution Rate % | 53.94 | #4 / 18 | Claude Code - Opus 5 (xhigh) | xhigh | Claude Code | 5 | 2026-09-21 |
| Terminal-Bench 4.0 (AA) | Resolution Rate % | 48.99 | #6 / 53 | Artificial Analysis - Claude Opus 5 (max) | Max | Artificial Analysis | 5 | 2026-10-01 |
| Code Arena | WebDev Overall | Arena Score | 1,694 | #6 / 54 | claude-opus-5-max | Max | — | 2 | 2026-09-30 |
| Agent Arena · Overall (Ranking / Pareto) | Net Improvement % | 8.76 | #5 / 39 | Claude Opus 5 (High) | High | — | 2 | 2026-09-30 |
| Artificial Analysis Coding Agent Index v1.5 | Coding Agent Index | 59.73 | #6 / 19 | Claude Code - Opus 5 (max) | Max | Claude Code | 1 | 2026-10-01 |
| DeepSWE v1.1 | Pass@1 % | 73.65 | #4 / 33 | mini-swe-agent - claude-opus-5 (max) | Max | mini-swe-agent | 5 | 2026-09-03 |
Every archived configuration
Artificial Analysis Intelligence Index v4.3 5 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Claude Opus 5 (max) | 50.7771 | — | — / Max / — | — / — | 2026-10-01 | Source · Archive |
| Claude Opus 5 (xhigh) | 49.6774 | — | — / xhigh / — | — / — | 2026-10-01 | Source · Archive |
| Claude Opus 5 (high) | 48.1219 | — | — / High / — | — / — | 2026-10-01 | Source · Archive |
| Claude Opus 5 (medium) | 44.8253 | — | — / Medium / — | — / — | 2026-10-01 | Source · Archive |
| Claude Opus 5 (low) | 39.3543 | — | — / Low / — | — / — | 2026-10-01 | Source · Archive |
Terminal-Bench 4.0 5 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Claude Code - Opus 5 (xhigh)Official Terminal-Bench 4.0 leaderboard row; score belongs to this harness × model × effort configuration. | 53.94 | 50.77 – 57.11 | Claude Code / xhigh / — | — / — | 2026-10-01 | Source · Archive |
| Claude Code - Opus 5 (max)Official Terminal-Bench 4.0 leaderboard row; score belongs to this harness × model × effort configuration. | 51.82 | 48.43 – 55.21 | Claude Code / Max / — | — / — | 2026-10-01 | Source · Archive |
| Claude Code - Opus 5 (high)Official Terminal-Bench 4.0 leaderboard row; score belongs to this harness × model × effort configuration. | 50.3 | 46.57 – 54.03 | Claude Code / High / — | — / — | 2026-10-01 | Source · Archive |
| Claude Code - Opus 5 (medium)Official Terminal-Bench 4.0 leaderboard row; score belongs to this harness × model × effort configuration. | 44.85 | 41.02 – 48.68 | Claude Code / Medium / — | — / — | 2026-10-01 | Source · Archive |
| Claude Code - Opus 5 (low)Official Terminal-Bench 4.0 leaderboard row; score belongs to this harness × model × effort configuration. | 34.85 | 30.91 – 38.79 | Claude Code / Low / — | — / — | 2026-10-01 | Source · Archive |
Terminal-Bench 4.0 (AA) 5 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Artificial Analysis - Claude Opus 5 (max)AA's own-harness independent measurement of Terminal-Bench 4.0 (the terminalBench40 field in the models-board payload, fraction×100); scored separately from the official tbench.ai board. | 48.9899 | — | Artificial Analysis / Max / — | — / — | 2026-10-01 | Source · Archive |
| Artificial Analysis - Claude Opus 5 (xhigh)AA's own-harness independent measurement of Terminal-Bench 4.0 (the terminalBench40 field in the models-board payload, fraction×100); scored separately from the official tbench.ai board. | 46.4646 | — | Artificial Analysis / xhigh / — | — / — | 2026-10-01 | Source · Archive |
| Artificial Analysis - Claude Opus 5 (high)AA's own-harness independent measurement of Terminal-Bench 4.0 (the terminalBench40 field in the models-board payload, fraction×100); scored separately from the official tbench.ai board. | 45.9596 | — | Artificial Analysis / High / — | — / — | 2026-10-01 | Source · Archive |
| Artificial Analysis - Claude Opus 5 (medium)AA's own-harness independent measurement of Terminal-Bench 4.0 (the terminalBench40 field in the models-board payload, fraction×100); scored separately from the official tbench.ai board. | 34.3434 | — | Artificial Analysis / Medium / — | — / — | 2026-10-01 | Source · Archive |
| Artificial Analysis - Claude Opus 5 (low)AA's own-harness independent measurement of Terminal-Bench 4.0 (the terminalBench40 field in the models-board payload, fraction×100); scored separately from the official tbench.ai board. | 26.2626 | — | Artificial Analysis / Low / — | — / — | 2026-10-01 | Source · Archive |
Code Arena | WebDev Overall 2 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| claude-opus-5-max | 1,694 | 1,687.5 – 1,700.5 | — / Max / — | — / — | 2026-10-01 | Source · Archive |
| claude-opus-5-high | 1,660 | 1,654.3 – 1,665.7 | — / High / — | — / — | 2026-10-01 | Source · Archive |
Agent Arena · Overall (Ranking / Pareto) 2 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Claude Opus 5 (High) | 8.76 | 7.26 – 10.26 | — / High / — | — / $2.5 | 2026-10-01 | Source · Archive |
| Claude Opus 5 (Max) | 8.55 | 6.98 – 10.12 | — / Max / — | — / $3.59 | 2026-10-01 | Source · Archive |
Artificial Analysis Coding Agent Index v1.5 1 configuration
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Claude Code - Opus 5 (max)Score belongs to this named agent/harness × model configuration; it is not a harness-independent model measurement. indexScore stored as percent (raw fraction × 100). | 59.7264 | — | Claude Code / Max / — | $10.79 / — | 2026-10-01 | Source · Archive |
DeepSWE v1.1 5 configurations
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| mini-swe-agent - claude-opus-5 (max) | 73.6486 | 69.776 – 77.521 | mini-swe-agent / Max / — | $11.84 / $10.43 | 2026-09-12 | Source · Archive |
| mini-swe-agent - claude-opus-5 (xhigh) | 73.1544 | 70.093 – 76.215 | mini-swe-agent / xhigh / — | $9.072 / $7.929 | 2026-09-12 | Source · Archive |
| mini-swe-agent - claude-opus-5 (high) | 72.8285 | 70.883 – 74.774 | mini-swe-agent / High / — | $6.076 / $4.969 | 2026-09-12 | Source · Archive |
| mini-swe-agent - claude-opus-5 (medium) | 68.9038 | 67.731 – 70.076 | mini-swe-agent / Medium / — | $3.29 / $2.482 | 2026-09-12 | Source · Archive |
| mini-swe-agent - claude-opus-5 (low) | 58.1292 | 55.798 – 60.46 | mini-swe-agent / Low / — | $1.663 / $1.216 | 2026-09-12 | Source · Archive |
These are reference mappings, not benchmarks of a subscription or API channel. Harness and quota-measurement effort alignment are unverified. Source task costs are not subscription task costs.
- Exact served-model reference; quota-measurement effort is unverified.
- Exact served-model reference only; product harness and quota-measurement effort are unverified. Not a benchmark measurement of this subscription or API channel.
Claude Opus 5 official API pricing
| Cached input | Input | Output |
|---|---|---|
| $0.5 | $5 | $25 |
Frequently asked questions
How much does Claude Opus 5 really cost per million tokens?
Measured as monthly fee ÷ tokens you can actually use, Claude Opus 5 ranges from $0.00772 (Claude Pro) to $0.7663 (Claude Opus 5 API) per million tokens.
Which subscription gives the most Claude Opus 5 usage?
Claude Max 20x (9/14+) allows roughly 15.7B tokens of Claude Opus 5 per month for $200.
What is the official API price of Claude Opus 5?
Official API list price per million tokens: $0.5 cached input, $5 input, $25 output — about $0.735 per million tokens on a standard workload mix.
How does Claude Opus 5 score on independent leaderboards?
Claude Opus 5 has scores on 7 leaderboards; its strongest placement is 50.78 (Intelligence Index) on Artificial Analysis Intelligence Index v4.3, #6 of 60 models tracked.
Who makes Claude Opus 5, and where can I use it?
Claude Opus 5 (model ID claude-opus-5) is made by Anthropic and is available through Anthropic, Factory.
More Anthropic models
| Model | Lowest real $/MTok | Cheapest plan | Plans |
|---|---|---|---|
| Claude Haiku 5.5 | $0.00028 | Claude Max 5x (9/14+) | 4 |
| Claude Sonnet 5.5 | $0.00273 | Claude Max 5x (9/14+) | 6 |
| Claude Sonnet 5 | $0.00309 | Claude Pro | 4 |
| Claude Opus 5.5 | $0.00545 | Claude Max 5x (9/14+) | 9 |
| Claude Opus 4.8 | $0.00772 | Claude Pro | 3 |
| Claude Fable 5 | $0.1083 | Claude Max 5x (9/14+) | 4 |
| Claude Fable 5.1 | $0.1157 | Claude Max 20x (9/14+) | 4 |