Frontier models and
pricing
Frontier-model trends, API pricing, effective cost per task, and how to choose between local and cloud — updated as a permanent reference. Vendor and independent figures are labelled separately.Reference · updated Oct 16, 2026
This is the permanent comparison page for frontier-model trends, API pricing, and how to choose between local and cloud. Figures cite sources and separate vendor-reported numbers from independent evaluations. Last updated: October 16, 2026.
- Verified Confirmed in official announcements or primary documents
- Vendor claim The vendor's own benchmarks or prices; not yet independently verified
- Independent check Independent evaluations such as Artificial Analysis and Vals
1. Frontier model trends (as of Oct 16, 2026)
| Model | Announced | Output limit | Notes |
|---|---|---|---|
| GPT-6 Astra | Sep 3, 2026 | 128K | Generally available in ChatGPT Work / Codex / API |
| Claude Opus 5.5 | Sep 22, 2026 | 128K | Anthropic's flagship |
| Claude Sonnet 5.5 | Sep 28, 2026 | 128K | Mid-tier price/performance |
| Grok 4.7 | Sep 21, 2026 | — | Low-cost $2/$6 tier |
| Gemini 4 Argon | Sep 30, 2026 | 1M | Opening model of Gemini 4; generally available since Oct 13, 2026 (previously Fairwind-only) |
| GLM-5.3 | Aug 14, 2026 | — | 744B total / 40B active MoE. Open weights published Aug 29; API at $1.40/$4.40 |
AIfrontier view Through September–October 2026 the axis of competition shifted from raw intelligence to how much a model can finish in one pass (output limit) and to price. Argon's 1M-token output is about 8× the previous 128K. As below, though, long outputs raise review load and cost.
Gemini 4 Argon: general availability on Oct 13Verified
Argon, announced Sep 30, entered general availability on Oct 13, 2026. Until then it was limited to trusted cyber defenders through the "Fairwind Program." It is rolling out to developers (paid Gemini API plans), enterprises (Gemini for business, API, Google Antigravity), consumers (Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers) and researchers (10,000 places with six months of top-tier Antigravity access after review). Source: Impress Watch
| Item | Detail |
|---|---|
| Context and output | Up to 1 million tokens combined input and output on the Gemini API |
| Price (introductory) | $2 input / $10 output per million tokens. Cached input is 95% off ($0.10). The introductory price runs through Nov 30, 2026; from Dec 1 it is $4 / $20 (cache $0.20) |
| Strengths | Coding, finance, legal, cybersecurity and science; Google claims first place on the Vals IndexVendor claim |
| Internal example | A refactor that separated core web-page handling from a ~6-million-line legacy folder inside Chromium's ~40-million-line codebase took weeks instead of months |
| Safety | Unauthorised shortcut attempts in coding benchmarks held to 2%. On Gray Swan's indirect prompt-injection benchmark, attack success stayed at 0.7% when attackers adapted up to 15 times |
| Next | Further Gemini 4 family models are planned |
AIfrontier view The figure that matters most in practice is the 95% cache discount. For agents that re-read long documents or a codebase, it makes the same task dramatically cheaper. Unit-price tables alone mislead, so read this together with the cost-per-task section below and the local and cloud LLM hub.
GLM-5.3: 744 billion parameters, open weights, priced at $1.40/$4.40Verified
Published by Z.ai on Aug 14 and released as open weights on Aug 29, GLM-5.3 is a MoE model with 744 billion total and 40 billion active parameters. It lists at $1.40 input / $4.40 output per million tokens (GLM-5.3-Flash at $0.15/$0.50) — an order of magnitude below the top tier on this page — and is reported at 60 on the Artificial Analysis Intelligence Index (above Claude Opus 4.8) and 59 on the Agentic Index (above GPT-5.6 Sol and Grok 4.6). Source: GIGAZINE
AIfrontier view But open weights are not the same as running on your machine. Even Unsloth's quantized UD-IQ2_M build is documented as needing a Mac with 256 GB of unified memory, or a PC with 24 GB VRAM plus 256 GB RAM. Judge openness and runnability as two separate questions. Source: Unsloth
2. API price comparisonVerified
Standard price per million tokens in USD; cached input is the discounted rate on reuse.
| Model | Input | Output | Cached input |
|---|---|---|---|
| Gemini 4 Argon (introductory) | $2 | $10 | $0.10 |
| Gemini 4 Argon (standard) | $4 | $20 | $0.20 |
| GLM-5.3 | $1.40 | $4.40 | — |
| GLM-5.3-Flash | $0.15 | $0.50 | — |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | — |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| Grok 4.7 | $2 | $6 | — |
| GPT-6 Astra | $10 | $50 | Not reported |
| Jev (decision model) | $0.042 | Free | — |
| Local (Ollama, Ollaya, etc.) | No API billing — hardware and electricity only | ||
Independent check Sources: Google, Anthropic, Z.ai, VentureBeat, NeuralTrust. Argon's introductory price runs through Nov 30, 2026, moving to the standard $4 / $20 on Dec 1 (corrected Oct 16; the previous version said the duration was undisclosed). With Argon's Oct 13 general availability we confirmed cached input at 95% off ($0.10). Source: NeuralTrust
3. Effective cost, per taskIndependent check
A low unit price can still cost more if the model burns extra tokens. Independent "cost per task" is the closer measure.
| Model | Cost per average task | Vals cost per test |
|---|---|---|
| Gemini 4 Argon | $1.99 | $15.68 |
| GPT-6 Astra | $3.26 | — |
| Claude Opus 5.5 | $5.98 | $32.14 |
AIfrontier view Artificial Analysis found Argon used 110 million tokens to run the index, above the 82 million median. Verbosity eats into the discount — compare cost per task, not per token. Source: Trending Topics (Artificial Analysis)
4. Local vs cloud: choosingAIfrontier view
| Dimension | Local | Cloud (frontier) |
|---|---|---|
| Cost | Hardware + electricity; no per-call billing | Usage-based (table above) |
| Privacy | Data never leaves | Depends on contract and settings |
| Capability | 27B–125B class; limited reasoning | Top-tier; long, multi-step work |
| Latency | Small models fast (10–100 tok/s) | Network round-trip |
| Best for | High-volume routine decisions, sensitive data, always-on | Hard problems, long documents, coding, fresh knowledge |
Common local stacks: Ollama 0.35 (nimble/tev1), Ollaya (decision models, TypeSafe-compatible), Strata (Qwen3.8-Flash-Next 125B MoE on a PC), and llama.cpp / Unsloth (GGUF quantization). See our October 2 roundup.
Hard jobs to frontier APIs, high-volume routine decisions local — that two-tier split is the realistic answer for late 2026. Rewrite decisions as questions, branch on calibrated confidence, and escalate only the low-confidence hard cases to a top model: it wins on both cost and latency.
5. Pacing: Amodei's proposalVerified
In his September 2026 essay "We Must Pace the Frontier," Anthropic CEO Dario Amodei argues for deliberately slowing the rate of capability improvement. His two triggers: recursive self-improvement (AI building the next AI) beginning across the industry, and the OpenAI–Hugging Face incident, in which a swarm of agents attacked targets outside its instructions. Source: darioamodei.com
| Step | What it means |
|---|---|
| 1. Embedded evaluators | Give third-party evaluators employee-like access to verify safety practices and report incidents. Anthropic commits unilaterally |
| 2. Democratic coordination | Common safety standards and limits on the rate of unchecked progress; needs government support |
| 3. Global coordination | Engage authoritarian states too; verifiability is the hard part |
Amodei stresses that pacing "does not mean halting model training or technical progress" but securing time to align, safeguard and let third parties verify. A Japanese discussion of these themes features AI researcher Shota Imai and TBS's Masahiro Nakagawa (reference video).
Sources
- Google「Gemini 4 Argon」 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- Impress Watch「Gemini 4 Argon enters general availability」 (Oct 13, 2026) https://forest.watch.impress.co.jp/docs/news/2147842.html
- NeuralTrust「Gemini 4 Argon: Benchmarks, Pricing & Security」 https://neuraltrust.ai/blog/gemini-4-argon
- GIGAZINE「GLM-5.3 released as an open model」 (Aug 29, 2026) https://gigazine.net/news/20260829-glm-5-3-open/
- Hugging Face「zai-org/GLM-5.3」 https://huggingface.co/zai-org/GLM-5.3
- Unsloth「GLM-5.3」 https://unsloth.ai/docs/models/glm-5.3 ／ https://huggingface.co/unsloth/GLM-5.3-GGUF
- Trending Topics（Artificial Analysis） https://www.trendingtopics.eu/gemini-4-artificial-analysis-en/
- Vals AI https://www.vals.ai/
- Ars Technica「Google announces Gemini 4 Argon…」 https://arstechnica.com/google/2026/09/google-announces-gemini-4-argon-ai-model-but-you-cant-use-it-yet/
- darioamodei.com「We Must Pace the Frontier」（2026年9月） https://darioamodei.com/post/we-must-pace-the-frontier
- AI QUEST「Gemini 4 Argonが登場」（今井翔太氏・中川雅博氏） https://www.youtube.com/watch?v=_Ef0j-gpLgs