Model Reference Refresh — Data as of 27 September 2026
Purpose: source-backed dataset and editorial recommendations for updating the site's model tables/calculators beyond the current May/July 2026 baseline. All figures below are drawn from vendor-primary sources (pricing pages, model cards, official announcements) retrieved via live search; anything not independently confirmed on a vendor page is flagged [UNVERIFIED].
1. Proprietary frontier / efficient models
OpenAI
| Model | Context / max output | Input $/M | Cached input $/M | Output $/M | Notes | Source |
|---|---|---|---|---|---|---|
| GPT-6 Astra (flagship, short ctx) | — | $10.00 | $1.00 | $50.00 | Long-context (>200K, up to 272K threshold) tier: $20/$2/$75. Frontier reasoning/agent model, rolling out to ChatGPT + API + Azure + Bedrock. | OpenAI pricing, GPT-6 Astra launch |
| GPT-5.6 Sol | 1.05M in / 128K out | $4.00 (promo, list $2 base tier per pricing page short-context row) | $0.20–$0.40 | $20.00 | Promotional pricing confirmed through at least 21 Nov 2026; >272K input tokens billed 2x input/1.5x output. | GPT-5.6 Sol model page |
| GPT-5.6 Terra | — | $4.00 (Fast-mode table) | — | $24.00 | "Balanced everyday work" tier; cut 20% at July 30 2026 price change. | Fast mode pricing, price-performance post |
| GPT-5.6 Luna | — | $0.10–$0.20 | $0.01–$0.02 | $0.50–$0.75 | Cut ~80% on 30 Jul 2026; cheapest tier, used for high-volume tool-using workflows. | Fast mode pricing |
| gpt-oss-120b / gpt-oss-20b | n/a (self-hosted) | n/a | n/a | n/a | Open-weight, Apache 2.0, not served via OpenAI API; run on vLLM/Ollama/llama.cpp. Baseline reference for OpenAI's open tier — no newer open-weight OpenAI release confirmed as of this date. | gpt-oss help center, GitHub |
Editorial note: "Fast mode" replaced "Priority processing" on 30 Jul 2026 (same service_tier values still accepted). Regional/data-residency endpoints carry a 10% uplift for eligible post–5 Mar 2026 models.
Anthropic
| Model | Context / max output | Input $/M | 5m cache write | Output $/M | Notes | Source |
|---|---|---|---|---|---|---|
| Claude Fable 5 / Mythos 5 | 1M in / 128K out (300K on Batch beta) | $10 | $12.50 | $50 | Fable 5 = GA flagship (9 Jun 2026, briefly suspended 12 Jun–1 Jul); Mythos 5 = safety-gate-lifted variant, limited availability via Project Glasswing. Cache-hit price differs from Fable 5.1 (see below). | Claude Fable 5 & Mythos 5, pricing table |
| Claude Fable 5.1 / Mythos 5.1 | 1M / 128K | $10 | $12.50 | $50 | Successor point release; cache-hit price $0.25/M vs $1/M on Fable 5. | Anthropic pricing |
| Claude Opus 5 | 1M / 128K (200K unless [1m] suffix set — confirmed community bug report) |
$5 | $6.25 | $25 | Launched 24 Jul 2026; ~half Fable-class price at near-frontier coding/agentic quality. US-only-inference SKU +10%. | Opus 5 launch, list-price PDF |
| Claude Opus 5.5 | 1M / 128K | ~$4 (per Sonnet-5 doc comparison table) | — | ~$20 | [UNVERIFIED price row] — appears only in a comparison table on the Sonnet 5 docs page, not confirmed against a dedicated Opus 5.5 pricing page; treat as directional until corroborated. | Sonnet 5 overview comparison table |
| Claude Sonnet 5 | 1M / 128K | $2 | $2.50 | $10 | Launched 30 Jun 2026; default on Free/Pro; drop-in for Sonnet 4.6 with adaptive thinking always on and legacy sampling params now erroring. | Sonnet 5 launch, pricing |
| Claude Haiku 4.5 (still current small tier) | 200K / 64K | $1 | $1.25 | $5 | No Haiku 5 confirmed as of 27 Sep 2026. | Anthropic pricing |
Editorial note: Anthropic ships no open-weight models; all Claude 5-family models are proprietary/API-only, deployed via Claude API, Bedrock, Vertex AI, and Microsoft Foundry.
| Model | Context / max output | Input $/M | Cached input $/M | Output $/M | Notes | Source |
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro (Preview) | up to 200K then long-ctx tier | $2.00 → $4.00 (>200K) | $0.20 → $0.40 | $12.00 → $18.00 | Frontier Gemini 3 tier referenced as the reasoning ceiling other vendors benchmark against. | Agent Platform pricing |
| Gemini 3.8 Flash | 1,048,576 in / 65,536 out | $0.75 (intro, through 31 Dec 2026) | $0.075 | $3.75 | Standard pricing doubles to $1.50/$0.15/$7.50 on 1 Jan 2027. Non-global regions +10%. | Gemini API pricing, Agent Platform pricing |
| Gemini 3.7 / 3.6 Flash | 1,048,576 / 65,536 | $0.75 (intro) | $0.075 | $3.75 | Same intro pricing/window as 3.8 Flash; positioned as prior-gen workhorses. | DevTk pricing summary — third-party aggregator, cross-check against ai.google.dev before publishing |
| Gemini 3.5 Flash / Flash-Lite | 1,048,576 / 65,536 | $1.50 / $0.30 | $0.15 / $0.03 | $9.00 / $2.50 | Retained as mid/cheap tier alongside Gemini 2.5 line. | Gemini API pricing (.md) |
2. Open-weight families available by 27 Sep 2026
| Family | License | Params (active/total) | Context | Notes | Source |
|---|---|---|---|---|---|
| DeepSeek-V4-Pro / V4-Flash | Open weights (DeepSeek license, check redistribution terms on HF card before calling it "open-source") | 49B/1.6T (Pro), 13B/284B (Flash) | 1M native, default across DeepSeek services | Released 24 Apr 2026; hybrid CSA/HCA attention, Muon optimizer; positioned just behind Gemini 3.1 Pro on world knowledge, ahead of other open models. | DeepSeek-V4 preview, HF model card, HF blog |
| Qwen3.8-Flash-Next | Open-weight preview (verify exact license on model card — likely Apache 2.0/Qwen license, not confirmed here) | 6B active / 125B total | 262,144 native, extensible to 1M | Released 26 Aug 2026; previews Qwen4 hybrid-attention architecture (Gated DeltaNet + Qwen Sparse Attention); efficiency-focused, not a capability flagship. | Unite.AI coverage — secondary source; confirm on Qwen/HF model card before treating specs as final |
| Llama 4 Scout / Maverick | Llama 4 Community License (custom, not OSI open-source; commercial-use gated over ~700M MAU) | 17B active / 109B (Scout), 17B active / 400B (Maverick) | 10M (Scout), 1M (Maverick) | Still Meta's newest publicly released Llama generation as of this date — no confirmed Llama 5 found in this research pass. Flag for re-check: verify no newer Llama release before publishing. | License, Model card |
| gpt-oss-120b / gpt-oss-20b | Apache 2.0 (true open-source) | 5.1B/117B, 3.6B/21B | not specified above (128K typical per model card) | OpenAI's only open-weight release; self-hosted only, not via OpenAI API. | gpt-oss GitHub |
Open-weight vs open-source distinction (for site glossary): Only gpt-oss carries a true OSI-style permissive license (Apache 2.0) with no usage-based gating. Llama 4 is open-weight but license-gated (custom commercial terms, MAU threshold). DeepSeek-V4 and Qwen3.8-Flash-Next are open-weight preview/production releases; exact redistribution terms should be pulled from each model's Hugging Face license field before the site asserts "open-source."
3. Editorial recommendations for the site
- model-comparison.ts is already current (dated 18/09/2026) with GPT-6 Astra, GPT-5.6 portfolio, Claude Fable 5.1, and Gemini 3.8 Flash — no changes needed there beyond adding Claude Opus 5 / Sonnet 5 and an open-weight row (DeepSeek-V4) for portfolio completeness.
src/routes/glossary.tsxis stale: context-window and tiering examples still cite "GPT-5," "Claude Opus 4.5," "Gemini 3 Pro," "DeepSeek V3.2," and "Llama 4 Scout" as if current-generation. Update to GPT-6 Astra/GPT-5.6, Claude Opus 5/Sonnet 5, Gemini 3.1 Pro/3.8 Flash, and DeepSeek-V4, while keeping Llama 4 Scout only as the long-context open-weight example (still accurate — no confirmed Llama 5).src/routes/index.tsxtrend-library card referencestrends/2026-pricing-landscape.mdlisting GPT-5/Opus 4.5/Gemini 3/DeepSeek V3.2 — either refresh that source file's contents or point the card at this new file plus a rewritten landscape doc.src/routes/dashboard.tsxsample spend row already says "Claude Sonnet 5" — consistent with the verified launch; no change needed, but confirm any adjacent sample rows (GPT/Gemini) are equally current.- Add Claude Opus 5 and Sonnet 5 as their own comparison entries (currently only Fable 5.1 represents Anthropic) since Opus 5/Sonnet 5 sit at materially different price/latency points relevant to routing decisions.
- Add one open-weight row (recommend DeepSeek-V4-Flash, given verified 1M context and cost-efficiency claims) to the comparison table and workload adviser so self-hosted/open-weight routing shows up as an option, clearly labeled "self-hosted — no vendor token price; cost = compute."
- Flag, don't publish, the Claude Opus 5.5 price row and the Qwen3.8-Flash-Next license until corroborated on a primary pricing/model-card page — both currently rest on secondary/comparison-table evidence only.
- Add a glossary note on Fast mode vs Priority processing (renamed 30 Jul 2026) and on Anthropic's tokenizer change at Opus 4.7 (affects "tokens per word" assumptions used in any cost calculator).
Verification status legend
- Confirmed on vendor primary source (pricing page, model card, or official blog): default for all rows above unless noted.
- [UNVERIFIED]: Claude Opus 5.5 pricing, Qwen3.8-Flash-Next exact license terms, DeepSeek-V4 redistribution license terms, Gemini 3.7/3.6 Flash pricing as sourced from a third-party aggregator (devtk.ai) rather than ai.google.dev directly.