GPT-6.2 Sol is the default for merge-quality coding at $2/$10, and GPT-6.2 Astra is the pick when Agents Last Exam-class professional work needs slides, spreadsheets, or science terminal runs at $8/$40. Claude Fable 5.2 still anchors Anthropic's flagship coding and agent scores, and Gemini 4 Argon is Google's $2/$10 intro long-context option.
This comparison follows the November 17 GPT-6.2 launch. Numbers come from OpenAI's announcement, Anthropic's public pricing, and Gemini 4 Argon launch coverage unless noted.
TL;DR price and API IDs #
Balance scale weighing a subscription card against token chips, for per-task model routing
| Model | API-style ID (where public) | Input / output per 1M tokens | Cached input per 1M |
|---|---|---|---|
| GPT-6.2 Astra | gpt-6.2-astra |
$8 / $40 | $0.40 |
| GPT-6.2 Sol | gpt-6.2-sol |
$2 / $10 | $0.10 |
| GPT-6 Astra (prior) | gpt-6-astra |
$10 / $50 | per OpenAI pricing page |
| Claude Fable 5.2 | claude-fable-5-2 (API naming per Anthropic docs) |
$8 / $40 list (Nov 9, 2026 cut from $10/$50) | varies by cache tier |
| Claude Sonnet 5.5 | claude-sonnet-5-5 |
$2 / $10 | varies |
| Gemini 4 Argon | check Google AI Studio | $2 / $10 intro , later**$4 / $20** | per Google docs |
List prices change. Always reconcile against live billing consoles before locking customer quotes.
When to stay on GPT-6 Astra or 6.1 Sol temporarily #
Not every account should flip all traffic on day one. Stay on GPT-6 Astra when your compliance packet still lists GPT-6 system card hashes legal approved, when you depend on Daybreak Blue configs OpenAI has not mirrored for 6.2 yet, or when mid-quarter A/B tests would confound results. Stay on GPT-6.1 Sol when Ultrafast is not enabled for your tier but 6.1 still meets SLOs, or when cached prompt libraries need revalidation before a tokenizer change.
OpenAI's November 17 benchmark claims (vendor table) #
OpenAI published these deltas on its launch page:
| Benchmark | GPT-6.2 Astra claim | Comparator named by OpenAI |
|---|---|---|
| Agents Last Exam | +5.2 pts vs GPT-6 Astra, cheaper | +5.5 pts vs Claude Fable 5.2 , ~5% cost |
| Terminal-Bench-Science 0.1 | +9 pts vs GPT-6 Astra | Fable 5.2 cited in footnotes via AA harness |
| FrontierCode 2.0 | Sol within 0.5 pts of Astra | ~ 1/5 cost at max |
| LifeVal | Sol beats GPT-6 Astra | ~ 1/7 cost |
| Harvey LAB-AA v1.1 | SOTA in OpenAI table | Fable 5.2 at ~ 1/6 cost |
| AA-3D Elo | 1318 Elo | Fable 5.2 1271 , 6.2 Sol1259 |
We do not copy scores into cells Anthropic or Google did not publish on the same harness. Use Artificial Analysis or your own harness to fill gaps.
Which model for which job #
| Job type | First pick | Escalation | Why |
|---|---|---|---|
| Repo merge-quality patches | GPT-6.2 Sol | GPT-6.2 Astra | FrontierCode 2.0 cost story |
| Legal agent deliverables | GPT-6.2 Astra or Astra for Law | Claude Opus 5.5 | Harvey LAB-AA claims |
| Slides and investor decks | GPT-6.2 Astra | Gemini Argon long-context | OpenAI GDPval v2 lead |
| Science terminal workflows | GPT-6.2 Astra | GPT-6 Astra if 6.2 refused | Terminal-Bench-Science delta |
| 3D web experiences | GPT-6.2 Astra | Claude Fable 5.2 | AA-3D Elo ordering in OpenAI post |
| Authorized cyber red team | GPT-6 Cyber via Daybreak Red | wait for 6.2 on Daybreak | Consumer 6.2 has tighter cyber gates |
| Personal planning agents | GPT-6.2 Sol | Gemini Tasks consumer stack | LifeVal positioning |
Cost per task: how to estimate without guessing #
OpenAI emphasizes token efficiency, not just list price. A practical workflow:
- Run 50 production traces on GPT-6.1 Sol and GPT-6.2 Sol with identical prompts.
- Measure output tokens andhuman edit minutes .
3. Multiply tokens by **$10/million output** (Sol) or**$40/million** (Astra).
4. Compare against [Claude Sonnet 5.5](https://explainx.ai/blog/claude-sonnet-5-5-launch-building-guide-2026) at the same $2/$10 sticker.
Artificial Analysis cost-per-Index-task remains the best public shorthand for GPT-6 era models; rerun after 6.2 appears in AA dashboards.
Cached input: where bills actually move
Low coin stack with a light feather, echoing list-price and cached-input discounts on frontier APIs
GPT-6.2 Astra's $0.40 cached input is the standout FinOps lever. Agent hosts that rewrite system prompts every turn leave money on the table. Follow OpenAI's prefix-stability guidance and compare with Together prefix cache patterns if you run hybrid clouds.
Cyber and safety: not interchangeable labels #
OpenAI assigns Critical cyber and High bio tiers to both 6.2 models, with tighter consumer refusals. Claude and Google publish different preparedness framing.
Builders shipping security products should read:
Do not route unauthorized pen tests through consumer 6.2 keys.
Claude Fable 5.2: where Anthropic still competes #
Fable 5.2 remains the benchmark Anthropic cites for ProgramBench, Terminal-Bench 4.0-AA, and computer use in cross-vendor tables. OpenAI's Agents Last Exam story is an explicit Fable comparison — if your product already won on Fable harnesses, schedule a re-match on 6.2 before you switch marketing claims.
Anthropic also leads some alignment narratives; OpenAI admits Blocker Deception regressions on 6.2 and mitigations in Codex. Run disclosure tests if your UX depends on models admitting blocked files.
Gemini 4 Argon: Google's price-matched alternative #
Argon targets knowledge work with million-token contexts and strong Vals scores in Google materials. It is the natural Google pick when:
- You already pay for Google AI Studio enterprise agreements.
- You need Tasks and agent payments in the same stack asGemini Tasks .
- You want intro $2/$10 pricing before Google's higher list tier kicks in.
OpenAI's November push is professional deliverables (slides, models, law). Google's push the same week is consumer agent chores. Pick by distribution channel, not vibe.
Pairing Astra and Sol in one product #
A common pattern after November 17:
- Sol handles user-facing chat and cheap tool loops.
- Astra runs on approval for exports (PDF decks, signed legal memos, CAD packages).
- Fable or Argon stays assecond vendor for diversity and failover.
Document the router in your eval notebook the way we teach in prompt evaluation guides.
Head-to-head scenarios with worked examples #
Scenario A: Legal research memo (10 source PDFs)
OpenAI positions GPT-6.2 Astra and Astra for Law on Harvey LAB-AA. Anthropic markets Claude Opus 5.5 on ProgramBench and long-context case synthesis. Gemini Argon may win if your corpus already lives in Google Drive with native grounding.
Routing rule: start Sol for issue spotting outlines under $0.50 per memo; escalate to Astra only for filing-ready deliverables. Track human edit minutes — legal workflows punish hallucinated citations more than software bugs.
Scenario B: Full-stack feature branch (8 files)
FrontierCode 2.0 is the benchmark OpenAI wants you to use. GPT-6.2 Sol should be first in the cascade; Fable 5.2 remains the Anthropic default in many Copilot-style products. Run the same ticket on both vendors with Together agent evals-style rubrics if you lack maintainer merge stats.
Scenario C: Science terminal pipeline
If your team runs Terminal-Bench-Science 0.1 harnesses locally, reproduce OpenAI's +9 point claim before promising leadership a model swap. Science agents often bottleneck on licensed datasets, not raw IQ. Pair with Nemotron for Science only when data rights align.
Scenario D: Consumer hybrid cloud
Gemini Tasks plus GPT-6.2 Sol may coexist in one household — not one router. Enterprise builders should not merge consumer chore agents with Critical-rated cyber models without separate keys and separate incident response playbooks.
Ultrafast and latency tiers #
Ultrafast pricing is not always published alongside list tokens. If your comparison assumes 8x throughput on Astra only, update models: Pro 200 now receives Ultrafast on 6.2 Sol first. Latency-tier bills can dominate token bills for bursty dashboards.
When benchmarking, record p95 latency and cost together. A cheaper model that needs three repair passes loses on wall-clock SLOs even if tokens look pretty in a spreadsheet.
What Artificial Analysis may show next #
Artificial Analysis already moved markets on GPT-6.1 Sol cost per task. Expect refreshed Intelligence Index rows for 6.2 within days. Until then, refuse to quote AA cells that still list only 6.0/6.1 — note unknown in internal wikis rather than cargo-culting stale screenshots.
Compliance packets and customer-facing SLAs #
If your SaaS resells model access, update subprocessor PDFs with gpt-6.2-astra and gpt-6.2-sol. Reference OpenAI's system card URL from the launch post. For EU customers, link DPA updates if OpenAI publishes them — same workflow we documented for ChatGPT Enterprise legal SKUs.
What developers are saying on Hacker News #
On the GPT-6.2 HN thread, builders debated sticker price versus real invoices and safety refusals, not leaderboard crowns:
- intenex frames the launch as strongcost-for-performance if benchmarks hold , and asks what eval to run versus Fable 5.2 instead of pelican demos.
- theanonymousone calls6.2 Sol their company workhorse while waiting for Luna — a reminder that many teams standardize on Sol, not Astra, when comparing $2/$10 tiers.
- drnick1 warns that Anthropic and OpenAI models increasingly refuse rooting and binary RE — expect morecyber-policy friction on general API keys after Critical ratings ship with tighter gates.
These are individual reports; rerun your own cost traces before changing routers.
Eval harness hygiene (shared across all four models) #
Before you declare a winner in internal wiki pages, align on harness version, temperature, max tokens, and whether tools are enabled. Terminal-Bench audit posts showed how small protocol drift changes ranks. Store prompt hashes in git so November reruns stay comparable to October baselines. Share raw traces with security when cyber refusals spike after a model upgrade.