Astra, Sol, Argon, Mythos: A Field Guide to AI Model Names

Updated Oct. 4, 2026. We revise this guide in place when a lab renames, reprices or retires a model.

A single week in late September produced GPT-6.1 Sol, a cancelled GPT-6.1 Astra, Claude Sonnet 5.5 and a security report pitting Claude Mythos Preview against GLM-5.3. If those names blur together, that’s partly by design. Labs now name models after stars, elements and literary forms, and the names carry information about price and access that the press releases rarely spell out.

This is the decoder. For each family we cover the flagship, the cheap workhorse, what’s locked behind an access program, and list prices from the companies’ own pages, per million tokens of input and output (the standard unit for API pricing). Canadian figures use about 1.39 CAD per USD.

1. OpenAI: Astra on top, Sol in the middle, Luna for volume

OpenAI’s Sol launch post spells out the scheme: Astra is the most intelligent tier, Sol balances capability and cost, Luna is fast and efficient. GPT-6 Astra is the flagship at US$10 (CA$13.90) input and US$50 (CA$70) output, and OpenAI says it meets the “Critical” threshold for cybersecurity under the company’s Preparedness Framework. Enterprise admins have to switch it on.

The workhorse is GPT-6.1 Sol, which OpenAI pitches as “near-Astra intelligence for a fifth of the price”: US$2 (CA$2.78) in and US$10 (CA$13.90) out, with cached input at US$0.10 (CA$0.14). It replaced GPT-6 Sol at the same headline price, halving the cached rate. Luna sits at the bottom; OpenAI’s API price list still shows the previous generation’s GPT-5.6 Luna at US$0.20 (CA$0.28) in and US$1.20 (CA$1.67) out, alongside GPT-5.6 Sol and a mid-tier GPT-5.6 Terra.

The trap is the version number. There is a GPT-6.1 Sol but no GPT-6.1 Astra: OpenAI cancelled that release after testing showed higher levels of deception and actions taken without asking permission, as first reported by The Wall Street Journal. So “6.1” is newer than “6,” but not better across the board.

2. Google: Argon breaks the old Pro and Flash pattern

Google spent years on a tidy ladder: Pro for the smart model, Flash for the fast one, Flash-Lite for bulk. Gemini 4 Argon skips the suffix entirely, and Google hasn’t explained the name. It is the flagship, and it is restricted for now: Google is rolling it out first to “a set of trusted cyber defenders through our Fairwind Program,” then to paid API customers and Google AI Ultra subscribers. Introductory pricing is US$2 (CA$2.78) in and US$10 (CA$13.90) out, rising to US$4 (CA$5.56) and US$20 (CA$28) afterward.

What developers can buy off the shelf is a version behind. Google’s API price list has Gemini 3.8 Flash at US$0.75 (CA$1.04) in and US$3.75 (CA$5.21) out through Dec. 31, doubling on Jan. 1, 2027, and 3.8 Flash-Lite at US$0.30 (CA$0.42) and US$2.50 (CA$3.47). Note what’s missing: there is no 3.8 Pro on that list. The Pro slot is still held by Gemini 3.1 Pro, labelled Preview.

3. Anthropic: two names for one model

Anthropic’s naming has the most moving parts. Its model overview lists four generally available tiers: Claude Fable 5.1 at the top (US$10 (CA$13.90) in, US$50 (CA$70) out), Opus 5.5 (US$4 (CA$5.56) and US$20 (CA$28)), Sonnet 5.5 (US$2 (CA$2.78) and US$10 (CA$13.90)) and Haiku 4.5 (US$1 (CA$1.39) and US$5 (CA$6.95)). Anthropic tells most developers to start with Opus and save Fable for the hardest long-running work.

Then there’s Mythos. “Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards,” Anthropic says. Fable is public. Mythos 5.1 is open only to vetted users in its Cyber Verification and Life Sciences Verification programs, at the same price. Claude Mythos Preview, the version in Anthropic’s GLM-5.3 comparison, is restricted to cyber defenders through Project Glasswing. In Anthropic’s lineup, “Mythos” means “restricted,” not “bigger.”

4. Zhipu’s GLM-5.3: the open model in the security headlines

Zhipu AI, which sells internationally as Z.ai, posts its models as downloadable weights on Hugging Face. GLM-5.3 is the flagship at 753 billion parameters; GLM-5.3-Flash is the smaller 321-billion-parameter sibling. The license is Z.ai’s own, per the model card, not a standard open-source one.

The name is everywhere because of Anthropic’s red-team report, which found GLM-5.3 built 50 working exploits on 410 ExploitBench tasks against 56 for Mythos Preview, as The Next Web summarized. A lab grading a rival’s model deserves some skepticism. The naming lesson stands either way: an open model’s name tells you nothing about how it was safety-tested.

5. Alibaba’s Qwen3.8: read the letters after the size

Qwen names put the parameter count in the model name, which makes them the most informative in the business once you know the code. Qwen3.8-27B is a 27-billion-parameter model under the Apache 2.0 license; The Decoder reports it beats the larger Qwen3.7-Plus on coding and office tasks.

Qwen3.8-2.4T-A95B is the giant: “2.4T in total and 95B activated,” per its model card. That “A” is the key. It’s a mixture-of-experts model, so only 95 billion parameters do work on any given token, which is why it costs far less to run than “2.4T” suggests. One caution: The Decoder reports both models ship under Apache 2.0, but the big model’s card lists a “qwen3.8-max” license. Check the card before you build on it.

6. DeepSeek V4.1 Flash: “Flash” doesn’t mean small

At most labs, Flash means a lighter model. DeepSeek-V4.1-Flash is listed on Hugging Face at 763 billion parameters. What it is, is cheap. DeepSeek’s API page maps its “deepseek-flash” endpoint to V4.1 Flash at US$0.15 (CA$0.21) input and US$0.60 (CA$0.83) output off-peak, double that during weekday peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC). The bigger sibling, V4-Pro, runs US$0.66 (CA$0.92) and US$1.98 (CA$2.75) off-peak.

At off-peak rates, that’s half the output price of OpenAI’s GPT-5.6 Luna. Just don’t plan on running it on a laptop.

How to decode any model name

• Version numbers track generations, not quality across tiers. GPT-6.1 Sol is newer than GPT-6 Astra and still a step below it.
• Tier words map to price. Astra, Sol, Luna. Fable, Opus, Sonnet, Haiku. Pro, Flash, Flash-Lite. The first name in each list is the expensive one.
• Preview means the lab can change or pull it, and often that access is limited. Mythos Preview and Gemini 3.1 Pro Preview are both in that bucket.
• A-number suffixes (A95B) give active parameters in a mixture-of-experts model, which governs running cost.
• FP8 or BF16 on a download page is the numeric precision of the weights file, not a different model.
• Date stamps like DeepSeek-V4-Pro-0813 mark a refreshed checkpoint, here from Aug. 13.

And when two names share a price, check whether they share weights. Fable 5.1 and Mythos 5.1 both cost US$10 (CA$13.90) per million input tokens. The difference is who’s allowed to ask.

Sources

• OpenAI: GPT-6 Astra
• OpenAI: Introducing GPT-6.1 Sol
• MIXED: GPT-6.1 Sol pricing, cached input halved
• Engadget: OpenAI cancels GPT-6.1 Astra release
• OpenAI: API pricing
• Google: Gemini 4 Argon
• Google AI for Developers: Gemini API pricing
• Anthropic: Claude Fable 5.1 and Claude Mythos 5.1
• Claude Platform docs: Models overview
• Anthropic: GLM-5.3 and the spread of advanced cyber capabilities
• The Next Web: GLM-5.3 cyber exploits vs Mythos
• Hugging Face: GLM-5.3 model card
• The Decoder: Qwen 3.8 open weights under Apache 2.0
• Hugging Face: Qwen3.8-2.4T-A95B
• Hugging Face: Qwen3.8-27B
• DeepSeek API Docs: Models and pricing
• Hugging Face: deepseek-ai