How to Read AI Model Versions
What a model name tells you, and what it doesn't
① Read a model name in four parts: company, tier, version number, suffix. A bigger number doesn't mean better in every way, and numbers from different companies can't be compared.
② The same name can behave differently depending on whether it is an alias or a pinned version. When a version changes, price, token counting, supported features and tone can change with it.
③ Models have lifespans. A model built into your workflow can disappear on a set date, so when a new one arrives, re-test on your own tasks rather than trusting the name.
① Getting started — new names keep arriving
AI news is a steady stream of new model names. Numbers go up, and tags like "mini," "Flash" or "preview" get attached. If you can read a name, it's easier to judge what a "new model launch" headline means for you. This piece uses one company's (Anthropic's) official documentation as the concrete example. Other companies' naming differs in detail but reads through the same framework. Examples are as of October 2026 and may change.
② A name has four parts
| Part | Example | What it tells you |
|---|---|---|
| ① Company / family | Claude, GPT, Gemini | Who built it. Even within a family, structure differs by generation |
| ② Tier | Fable · Opus · Sonnet · Haiku (other companies: Pro/Flash, mini, etc.) | A split by size and purpose. Larger models are stronger at hard reasoning and slower and costlier. Smaller ones are faster and cheaper |
| ③ Generation / version number | 5, 5.1, 5.5 | The order of release within a family. A decimal is an improved release, but doesn't denote a "half step" of any measurable size |
| ④ Suffix | A date (20250929), preview, latest, thinking | Whether it's pinned, a trial release, or a continually changing alias. Terms vary by service |
In Anthropic's October 2026 documentation, the current models split into four tiers. Fable 5.1 is described as for "demanding reasoning and long-horizon agentic work," Opus 5.5 for "long-running agentic coding and knowledge work," Sonnet 5.5 as "the best combination of speed and intelligence," and Haiku 5.5 for "high-volume, latency-sensitive tasks such as classification, extraction, and routing." A tier name is the company's guidance on what the model is suited for.
③ What a number doesn't tell you
- Numbers across companies aren't comparable. The 5 in "GPT-5" and the 5 in "Claude 5" are not the same scale. Each company numbers its own family's generations.
- The higher number isn't better at everything. Even within a generation, price and speed differ widely by tier, as the table below shows. Improvements across generations also vary by area. For cautions on comparing scores, see Can You Trust Benchmark Scores?.
- A smaller model is enough for many jobs. For fixed-format work such as classification, extraction and summarization, a small tier is faster and cheaper. Reserving the large model for tasks that need hard reasoning is the cost-sensible split.
| Tier (Oct. 2026) | Price per 1M tokens (input / output) | Relative speed | Suited for (company guidance) |
|---|---|---|---|
| Fable 5.1 | $10 / $50 | Slower | Demanding reasoning, long-horizon agentic work |
| Opus 5.5 | $4 / $20 | Moderate | Long-running coding, knowledge work |
| Sonnet 5.5 | $2 / $10 | Fast | Balance of speed and intelligence |
| Haiku 5.5 | From $0.10 / from $0.50 | Fastest | High-volume, low-latency classification, extraction, routing |
Source: Anthropic Models overview. Prices are base API rates; batch and caching discounts are separate.
④ Same name, different content — aliases and pinned versions
Sometimes the model a name points to changes even though the name stays the same. Distinguish two kinds.
- Pinned version (snapshot): carries an identifier such as a date and stays that model. Anthropic states that from the 4.6 generation on, even dateless names (for example
claude-sonnet-5-5) are themselves pinned versions. - Alias: a convenience name pointing to the newest model. In earlier generations, the alias pointed to a dated, pinned version. Depending on the service, what an alias points to can change.
A chat app's "default model" or "auto-select" setting carries the same issue. If yesterday's and today's answers differ, it may be the model behind the scenes that changed rather than your question. If you've built AI into a workflow, it's worth recording the model name and version you use.
⑤ What changes along with a version
A new version isn't just "smarter." If you use it for work, all of the following move with it.
| What changes | What happens |
|---|---|
| Price | Unit price differs by generation and tier. Moving the same task to a pricier tier can raise cost a lot |
| Token counting (tokenizer) | According to Anthropic's documentation, the tokenizer introduced with Opus 4.7 gives about 555k words per 1M tokens, versus about 750k words for earlier models. The same text is counted as more tokens, so limits and cost change even for the same text. For the mechanics, see Tokens and the Context Window |
| Supported features and settings | For example, settings such as temperature, top_p and top_k were removed in API version 2026-10-21, and supplying them returns an error (400). Switching versions can break existing code |
| Knowledge cutoff | The point where the model's knowledge ends shifts. Current models list June 2026. Without a search tool, they don't know what happened after that |
| Tone and output style | Answer length, formatting and how instructions are followed change. The documentation also keeps separate prompting guidance per model, so there is no guarantee an old prompt gives the same quality |
⑥ Models have lifespans too
Anthropic separates the model lifecycle into Active → Legacy → Deprecated → Retired. Requests to a retired model fail. The documentation says retirements of publicly released models are announced at least 60 days ahead (Anthropic Model deprecations).
- Sonnet 4 and Opus 4 were deprecated on April 14, 2026 and retired June 15.
- Sonnet 4.5 (
claude-sonnet-4-5-20250929) was deprecated on September 30, 2026 and is scheduled to retire November 30. The recommended replacement is Sonnet 5.5. - Partner platforms such as Amazon Bedrock set their own schedules, so dates for the same model can differ.
An automation set up once and forgotten can stop suddenly on its retirement date. Keep a list of the model names you use and check for deprecation notices. When migrating, question 7 in What to Check Before Adopting AI at Work, "Is there a way out?", is the one that actually gets used.
⑦ What to check when a new model arrives
- Look at company and tier first. Is it a large or small model, and is the tier right for your use?
- Compare numbers only within one company. Check in a table how price, speed and limits differ from the previous generation.
- Check whether it's an alias or a pinned version. If the name can keep changing, see whether you can pin it before relying on it.
- Recalculate tokens and cost. The same document can become more tokens.
- Check supported settings and the knowledge cutoff. Test whether existing code and prompts still work.
- Compare on five of your own tasks. Run them under the same conditions rather than trusting a score chart. The method is in the "30-minute self-test" in Comparing AI Tools.
- Write down the retirement date. Check periodically whether a deprecation notice has appeared for the model you use.
⑧ Investor angle — how to read a model launch
A new model launch is used as an event for related stocks. But as the tables above show, a name change connects to three numbers: unit price (which tier is used), token consumption (tokenizer and reasoning style), and replacement cycle (retirement schedule). The larger these are, the more demand for servers, chips and memory. The big picture of AI infrastructure demand is in the AI Semiconductor Sector Analysis, and where money flows among model builders is in the AI Circular Financing Analysis. A new model name alone is not earnings, though; whether demand actually shows up is confirmed in the numbers of an earnings report.
⑨ Practical checklist
□ Did you split it into company, tier, version number and suffix?
□ Did you compare numbers only within the same company's family?
□ Did you ask whether this task really needs a large model or a small one would do?
□ Did you check whether it is an alias or a pinned version?
□ Did you recalculate token count and cost for the same document?
□ Did you test whether existing prompts, code and settings still work on the new version?
□ Did you record the model name you use and its retirement schedule?
□ Did you compare on a few of your own tasks instead of a score?
※ Model names, prices and dates here use Anthropic's official documentation as of October 2026 as an example and may change. Rules at other companies should be checked against their own official documentation. This is not an assessment of any product, and investment-related content is not investment advice. Investment decisions and their consequences are the investor's own responsibility.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.