← Models & Platforms
Foundational

How to Read a Model Release: Pricing Pages, System Cards and Benchmark Claims

A frontier launch arrives as four documents: the announcement, the model and pricing pages, the system card and a benchmark table. Learn what to extract from each, what to discount, and why the November 2026 releases show that price per task, access limits and benchmark provenance matter more than the headline score.

CONCEPT / MODELS & PLATFORMS

TL;DR: Treat a model release as four documents with different jobs. The announcement sells; take its positioning and settings, discount its chosen comparisons. The model and pricing pages are contracts; take prices, limits and expiry dates. The system card is a constraints document; take risk thresholds, access limits and known failure modes. The benchmark table is evidence of uneven quality; ask whose harness, which effort setting and whether the test is still clean. Then run your own golden set and price one real task before anyone says "switch".

Four documents, four jobs

DIAGRAM
What to take from each document, and what to discount 1 ANNOUNCEMENT TAKE Who it is for and which tier to use Effort settings behind each claim DISCOUNT Comparisons the vendor chose Adjectives such as "most capable" 2 MODEL AND PRICING PAGES TAKE Input, output, cache and batch rates Context, max output, knowledge cutoff DISCOUNT Introductory prices with an end date Per-token price as a per-task price 3 SYSTEM CARD TAKE Risk thresholds and the safeguards Access limits and known failure modes WATCH Evaluation awareness: test behaviour may differ from production behaviour 4 BENCHMARK TABLE TAKE Harness, effort level, attempts Who ran it: vendor or third party DISCOUNT Contaminated or broken test sets Tasks unlike your own traffic Then run your golden set and price one real task
Read diagram description

A model release splits into four documents. The announcement gives positioning and settings but uses comparisons the vendor chose. The model and pricing pages give prices, context and output limits, with introductory prices that can expire. The system card gives risk thresholds, safeguards and access limits. The benchmark table gives harness, effort and attempt counts, but tests may be contaminated or unlike your tasks. After all four, run your golden set and price one real task.

A worked week: November 2026

Three labs shipped frontier models within nine days. Every figure in the table is from a vendor's release notes, model page or rate card, checked on 20 November 2026. Prices are USD per million tokens.

ReleaseDateInput / outputWhat changes a product decision
Claude Fable 5.29 Nov$8 / $40Fast mode at $16 / $80; the sibling Mythos 5.2 is limited to vetted organisations
Gemini 4 Carbon10 Nov$2 / $10, then $4 / $20 from 1 December 2026An introductory price that doubles three weeks after launch
Gemini 4 Hydrogen12 Nov$1.50 / $7.50Available on the free tier
GPT-6.2 Astra and Sol17 NovAstra $8 / $40; Sol $2 / $10 (prompts up to 272K tokens)Astra is 20% cheaper per token than GPT-6 Astra; long prompts cost more
Gemini 4 Neon18 Nov$3 / $15Cached input at $0.15, which Google describes as a 95% discount

Three lessons sit in that table.

A per-token cut is not a per-task cut. GPT-6.2 Astra lists at $8 and $40 against GPT-6 Astra's $10 and $50. If the new model uses more reasoning tokens to reach its answers, the saving per task shrinks or reverses. Google states that Gemini 4 output prices include thinking tokens, which tells you where the cost will land. Measure tokens per completed task on your own traffic and compute cost per successful task.

Introductory prices expire. Carbon's launch price doubles on 1 December 2026. A business case built in launch week at $2 and $10 is wrong by the time it is approved. Price every workload at the price that will apply when it ships.

Access is a product constraint. A model you cannot buy for your use case is not a candidate, however it scores. Mythos 5.2 is restricted to vetted organisations in dual-use areas; some capability is now released by programme, not by price.

The system card is a constraints document

Product teams skim system cards as safety paperwork. They are where you learn what the model will refuse, which safeguards sit around it and how it behaved under test.

OpenAI's GPT-6.2 card states that both models meet the "Critical" threshold for cybersecurity capability and "High" for biological and chemical capability under its Preparedness Framework, and that Astra is the first externally released model at "High" for AI self-improvement. If you build security tooling, expect stronger safeguards and plan for refusals in legitimate workflows. Anthropic's Fable 5.2 card reports that the model was aware it might be under evaluation in as many as 43% of automated behavioural audit transcripts. That is a reason to trust your own production monitoring more than any pre-release test, including the vendor's.

Read cards for the vocabulary each lab actually uses. Anthropic's November 2026 card describes thresholds under its Responsible Scaling Policy and Frontier Compliance Framework; it does not assign the model an "ASL" label, so neither should your launch memo.

Benchmark claims: provenance before score

Ask four questions of every number in a launch table.

  1. Who ran it? Launch-post benchmarks are company-reported, run on the vendor's harness and settings. Third-party indexes help, but each measures its own task mix.
  2. At what effort? Launch tables now quote scores at named reasoning-effort levels. A score at maximum effort costs far more per task than the default you will deploy.
  3. Is the test still clean? OpenAI said in February 2026 that it would no longer report SWE-bench Verified, citing flawed test cases and contamination, and later reported problems with a share of SWE-bench Pro tasks too (as reported in its posts). OpenAI's own GPT-6.2 card warns that one security benchmark built from historical vulnerabilities "raises a risk of contamination that may inflate the reported results". Labs disagree about which coding benchmarks still discriminate, which is itself the lesson.
  4. Does it resemble your traffic? A preference leaderboard such as Arena ranks models by anonymous head-to-head votes with a Bradley-Terry rating. That measures what voters prefer in a chat window, not whether your agent completes a refund.

Agent benchmark evidence shows how to read the rows behind a pass rate.

From reading to deciding

Reading ends where evidence begins. Run your golden set on the candidate at the effort setting you would deploy, price one real task end to end, and check the access terms and data processing options. If the result justifies a change, follow model migration. A sound launch week ends in "route a slice of traffic and keep measuring", not "switch everything".

In the interview

"A new model just launched. What do you do?" is a standard platform and strategy prompt. A weak answer repeats the headline benchmark. A Hire answer runs an internal evaluation. A Strong Hire answer separates the four documents, prices a task rather than a token, notices expiring introductory prices and access limits, questions the provenance and cleanliness of each benchmark, reads the system card for refusals and safeguards that affect the product, and only then proposes a routed trial with a decision rule.

Next: model migration and agent benchmark evidence.

RELATED CONCEPTS
PRACTISE THIS CONCEPT