
How many Claude models are there? As of November 17, 2026, twelve Claude versions are on sale here, spread across four lines: Fable, Opus, Sonnet and Haiku. The answer is not "buy the newest and biggest". It is three lines: use claude-sonnet-5-5 for everyday development ($0.099 / $0.495 effective, $4.46 for a thousand requests a month); use claude-haiku-5-5 for high-frequency lightweight work ($0.0088 / $0.044, $0.40 a month, the cheapest in the family); and switch to claude-opus-5-5 for hard refactors and long-running agents ($0.198 / $0.99, $8.91 a month). The one to avoid is claude-opus-4-6: the oldest Opus, and the priciest, $49.50 a month — 5.6 times opus-5-5.
A Claude name carries three pieces of information. The product line sets the positioning, the number only tells you which generation it is, and the seller's discount decides what you actually pay. Mix the three up and you will conclude, wrongly, that newer always costs more.
Fable is the Mythos tier, built for autonomous knowledge work and coding, and the top of the four. Opus is the flagship, aimed at hard reasoning, end-to-end software tasks and code review. Sonnet is the balanced workhorse for coding and agents. Haiku is small and fast, for summaries, subagents and browser automation.
Watch the numbers too: current versions run to 5.5, while 4.5, 4.6, 4.7 and 4.8 are still listed. They are not defective, but their discount structure has nothing in common with the newest releases. There is a section on that below.
Sorted by effective input price, in USD per million tokens. "Discount" is what you pay relative to list — 0.5 means 5% of the official rate.
| Model | Official (in/out) | Discount | Effective (in/out) | Context | Sellers |
|---|---|---|---|---|---|
| claude-haiku-5-5 | $0.1 / $0.5 | 91.2% off | $0.0088 / $0.044 | 1M | 3 |
| claude-haiku-4-5 | $1 / $5 | 95% off | $0.0495 / $0.2475 | 200K | 1 |
| claude-sonnet-5-5 | $2 / $10 | 95% off | $0.099 / $0.495 | 1M | 5 |
| claude-sonnet-4-6 | $3 / $15 | 95% off | $0.1485 / $0.7425 | 1M | 1 |
| claude-sonnet-5 | $2 / $10 | 91.2% off | $0.176 / $0.88 | 1M | 3 |
| claude-opus-5-5 | $4 / $20 | 95% off | $0.198 / $0.99 | 1M | 5 |
| claude-opus-4-8 | $5 / $25 | 95.6% off | $0.22 / $1.1 | 1M | 3 |
| claude-opus-4-7 | $5 / $25 | 95% off | $0.2475 / $1.2375 | 1M | 2 |
| claude-opus-5 | $5 / $25 | 91.2% off | $0.44 / $2.2 | 1M | 3 |
| claude-opus-4-6 | $5 / $25 | 78% off | $1.1 / $5.5 | 1M | 1 |
| claude-fable-5 | $10 / $50 | 78.55% off | $2.145 / $10.725 | 1M | 3 |
| claude-fable-5-1 | $10 / $50 | 78.55% off | $2.145 / $10.725 | 1M | 3 |
The spread between cheapest and priciest is 244x: fable-5-1's effective input price is $2.145, haiku-5-5's is $0.0088, and both are first-party models. That gap comes from list price ($10 versus $0.1) multiplied by discount (78.55% off versus 91.2% off), not from a capability difference of that size.
Start with a chart. This is the effective monthly cost of the five Opus versions under one usage profile — 30 million input and 3 million output tokens a month:
The result is blunt: the newest, opus-5-5, is the cheapest here at $8.91; the oldest, opus-4-6, is the most expensive at $49.50 — a 5.6x gap. There is exactly one reason: opus-4-6 is only quoted at 78% off, while opus-5-5 reaches 95% off. Same work, five times the money, zero extra capability.
The same inversion holds on the Sonnet line: sonnet-5-5 ($4.46) is 33% cheaper than sonnet-4-6 ($6.68) and 44% cheaper than sonnet-5 ($7.92). When the new version is quoted at 95% off and the older one at 91.2%, the effective prices simply flip.
Supply explains it. opus-5-5 and sonnet-5-5 each have five sellers quoting (two of them certified), and that competition drives the best quote to 95% off with the richest ladder — five tiers from 95% to 50% off. opus-4-6 and sonnet-4-6 have a single seller and a single tier, so there is nothing to negotiate against.
Real traffic through this platform's gateway in the last 24 hours (data as of November 17, 2026): opus-5-5 saw 32 calls, 100% success, 11.6-second P50 latency, 77.7 tokens per second out. sonnet-5-5 saw 19 calls, 100% success, 14.9-second P50, 72.9 tokens per second. fable-5-1 saw 27 calls, 100% success, a 5.6-second P50 to first token, but only 4.14 tokens per second of output.
Read that with restraint. Twenty or thirty calls prove these three paths work and stay up; they do not prove one is a tier faster than another. fable-5-1's 4.14 tokens per second especially should not be read as "slow": first token in 5.6 seconds means fast reasoning followed by a deliberately slow stream, the signature of a heavy-thinking model rather than a stalled one.
The other nine versions saw one call or none in 24 hours, so no latency or success figure exists. Saying so plainly: for those, prices are comparable, performance is not.
Everyday features and bug fixes (the default): claude-sonnet-5-5. Five sellers, 95% off, $4.46 a month, 1M context, 128K max output, 19 measured calls with zero failures. The least controversial value in the family.
High-frequency lightweight work (classification, summaries, subagents, page scraping): claude-haiku-5-5. At $0.0088 / $0.044 effective and $0.40 a month, it is 5.6 times cheaper than the previous haiku-4-5 ($2.23) — and its context grows from 200K to 1M while max output doubles from 64K to 128K. Using the old Haiku for high-volume work is simply burning money.
Hard problems (large refactors, cross-file changes, code review): claude-opus-5-5. Its own list price is already $4/$20, one tier below opus-5 at $5/$25, and 95% off lands it at $0.198 / $0.99 — less than half of opus-5 ($0.44 / $2.2). Thirty-two measured calls, zero failures, 77.7 tokens per second: the fastest Claude in this batch.
Only for specific cases: claude-fable-5-1. It is the Mythos tier, aimed at autonomous knowledge work and long workflows, at $2.145 / $10.725 effective — 21.7 times sonnet-5-5. One improvement deserves a mention: fable-5-1 cut cache reads to a quarter of fable-5's price ($0.25 versus $1.0 per million tokens). If your workload asks repeated questions against one very long document, the cache-read share is high and that difference recovers real money. Otherwise there is no case for paying 21 times more.
Cache writes and reads. Every Claude model supports prompt caching. Cache writes are priced on a 5-minute TTL, and the 1-hour TTL costs more (Opus at $10 per million tokens, Sonnet at $4, Fable 5 at $20). If the same context gets asked twice or more within five minutes, cache reads usually cost a tenth of the input rate — a very visible saving.
Batch runs at half price. Input and output are both 50% off on the Batch API across the family. Evaluation sweeps, bulk copy rewrites and mass test generation are not urgent, so route them through Batch and the effective rate drops again — sonnet-5-5's $0.099 becomes $0.0495.
1M context at no surcharge. Everything except Sonnet 4.5 bills its full 1M window at the standard rate, with no long-context premium. Gemini 3.1 Pro, by contrast, doubles its rate past 200K tokens. That matters when you are chewing through a large repository.
Pick by scenario, then check the table: lightweight work goes straight to haiku-5-5, never haiku-4-5; daily development locks onto sonnet-5-5; hard problems switch to opus-5-5. Run all three and, under the basis above, the combined monthly bill is $0.40 + $4.46 + $8.91 — still in the low teens.
Prices and seller discounts move, so the numbers here describe November 17, 2026 only. Before you order, check live quotes and seller counts on the model marketplace, and verify context window, max output and cache terms for each version on the model docs — when old and new versions get mixed, those parameters diverge even more than the prices do.
One line to close: of these twelve Claude versions, every wrong pick comes from sorting by version number instead of by effective price.
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa
Hailuo 2.3-fast has the lowest effective price (about $0.00185/s), Kling 3.0 offers the fullest range up to 4K, and Seedance 2.5 costs most but render