Frontier Intelligence Directory · Latest registry change October 9, 2026
Explore selected providers and models by capability, inspect their sources, and compare candidate routes. Live pricing is available for mapped models; curated entries retain their source context.
Connect provider sources and catalog prices to task-specific candidates, creator workflows and public receipts. Each recommendation still needs acceptance evidence from your workload.
10
Providers tracked
43
Frontier models
0
Agentic platforms
21
Live-priced
September release · Sources checked 7 September 2026
The new entry connects model facts, pricing and three free practical guides. Twelve synthetic evaluation cases are defined and have not run; treat any capability claim below as vendor-reported until a scored evaluation run backs it.
This update covers Astra and the Sonnet 5 price correction. The candidate policies below have their own review date. Historical comparisons retain their original scope.
Candidate policies reviewed 10 September 2026. Each names what must pass before adoption; none certifies a production route.
This historical example applies fixed editorial rules. It does not query receipts, execute models or validate a production route.
Select one or more constraints to trigger routing recommendation.
| If you need… | Pick | Runner-up | Why |
|---|---|---|---|
| Maximum capability; spend unconstrained | GPT-6 Astra | Claude Opus 5 | Compare independent solutions on the hardest real tasks. Accept only after evidence review; extra inference is useful only when it improves the accepted result. |
| Software and web development | Claude Sonnet 5 | GPT-6 Astra | Start with a bounded implementation candidate. Escalate unresolved failures; require build, regression tests, browser checks and a reviewable diff. |
| Game development | GPT-6 Astra | Claude Sonnet 5 | Evaluate engine integration, deterministic simulation tests, frame-time budget and playability separately. Code benchmarks cannot establish game quality. |
| Books and long-form writing | Claude Opus 5 | GPT-6 Astra | Compare voice fidelity, continuity, revision burden and reader preference against a locked chapter brief. Use a manuscript memory system and independent continuity checks. |
| Content creation and research | Claude Sonnet 5 | Gemini 3.7 Flash | Require traceable sources, brand review and editable deliverables. Judge the whole workflow; use specialist image, audio and video tools for media output. |
| Bulk extraction and support triage | Gemini 3.7 Flash | Claude Haiku 4.5 | Measure schema correctness, abstention, escalation and tail latency. Compare total spend per accepted item; preserve a human route for consequential exceptions. |
| Self-hosting and data control | Mistral Large 3 | gpt-oss (120b / 20b) | Inspect model license, deployment location, GPU capacity and operating cost. Open weights alone establish neither free inference nor compliance. |
| Long-running xAI agents | Grok 4.6 | Grok 4.3 | Compare completion, recovery from tool failures, context growth and spending limits on the same task. External benchmarks do not qualify the production route. |
Estimate standard short-context token spending. This example excludes retries, tools, cache creation and review time; it does not measure ROI or execution quality.
Illustrative token allocation
Illustrative split: 80% of input and output tokens to Fast-Path (Gemini 3.8 Flash) + 20% to Deep-Reason (Claude Opus 5.5). Saves $52.00/mo (65%) vs 100% Claude Opus 5.5.
Anthropic
Registry · USD per million tokens. Standard Claude API rates in USD for prompts up to 100,000 tokens. A request whose prompt is over 100,000 tokens pays $0.50 / $2.50 per million input/output tokens for the full request. Cache writes, batch, geography and partner-platform modifiers excluded.
DeepSeek
Registry · USD per million tokens. Off-peak DeepSeek API rates (cache-miss input, output). Peak rates are double: input $1.32, output $3.96 per million tokens, 01:00-04:00 and 06:00-10:00 UTC Monday to Friday, excluding Chinese public holidays. Cache-hit input is $0.022 off-peak and $0.044 peak.
Off-peak rates. Peak rates are 2× (DeepSeek pricing page).
Google DeepMind
Registry · USD per million tokens. Standard Gemini API introductory rates in USD through 2026-12-31. From 2027-01-01: input $1.50, output $7.50, cached input $0.15 per million tokens. Cache storage and tools charged separately.
xAI
Registry · USD per million tokens. xAI API rates in USD for prompts under 200K tokens; a request with 200K prompt tokens or more pays $4 / $12 per million input/output tokens ($1 cached input). Tool calls, the US regional endpoint premium and the Cursor/Grok Build fast variant are excluded.
Anthropic
Registry · USD per million tokens. Standard Claude API rates in USD; cache writes, batch, fast mode, geography and partner-platform modifiers excluded.
OpenAI
Registry · USD per million tokens. Standard USD rates for prompts up to 272K input tokens. A request with more than 272K input tokens pays $4 / $15 per million input/output tokens ($0.20 cached input) for the full request. Cache writes, tools and other processing tiers are charged separately.
Anthropic
Registry · USD per million tokens. Standard Claude API rates in USD; cache writes, batch, fast mode, geography and partner-platform modifiers excluded.
Anthropic
Registry · USD per million tokens. Standard Claude API rates in USD; cache writes, batch, fast mode, geography and partner-platform modifiers excluded.
OpenAI
Registry · USD per million tokens. Standard USD rates for prompts up to 272K input tokens. Above 272K, the full request uses 2x input/cache and 1.5x output rates. Cache writes, tools and other processing tiers are charged separately.
Pick the job, jump to the providers that lead.
Complex problem-solving, math, abstract reasoning, long-horizon planning
Vision, document, chart, and cross-modal reasoning across text/image/audio
Tool use, function calling, agent SDKs, computer use, long-horizon execution
Native speech in/out, real-time conversation, audio understanding
Sort and filter every tracked model. Live pricing via OpenRouter where available. Click a model for the full breakdown.
43 models live pricing via OpenRouter
Creators don’t pick a model — they assemble a stack. Here’s what to use across each modality, and the workflow.
Generate, arrange, and master AI music
Workflow. Draft lyrics + style prompt with Claude → generate in Suno → iterate sections → export and master. The LLM is the creative director; Suno is the band.
AI Music MasterclassGenerate and edit on-brand visuals
Workflow. Concept + prompt with an LLM → generate hero with Imagen/Nano Banana → object-level edits (swap, resize, recolor) → export to the design system.
Visual creation systemGenerate and edit video with natural language
Workflow. Script with an LLM → generate with Omni → edit by instruction (background swap, camera angle) → produce. Increasingly a single agentic sequence.
Long-form, on-voice writing at depth
Workflow. Research + outline with Opus (1M context holds your whole corpus) → draft → tighten with Sonnet → publish. The brand voice gate stays human.
Content StudioShip code with agentic assistants
Workflow. Flash for the high-volume agent loop, Opus 4.6 for the critical reasoning path. Run inside Claude Code, Cursor, or Antigravity 2.0.
Frontier Models ArenaFlagship model, capability focus, agentic platforms, and notable tech for every tracked provider.
Flagship
Claude Fable 5.1
Context
1M
Price (in/out)
$10/$50
Released
2026-09-01
Models (13)
Flagship
GPT-6 Astra
Context
1.05M
Price (in/out)
$10/$50
Released
2026-09-04
Models (8)
Flagship
Gemini 3.8 Flash
Context
1.05M
Price (in/out)
$0.75/$3.75
Released
2026-09-02
Models (7)
Flagship
Grok 4.7
Context
500K
Price (in/out)
$2/$6
Released
2026-09-21
Models (1)
Models (2)
Models (2)
Models (1)
Models (1)
Where the models actually do work — IDEs, CLIs, desktops, agent platforms, managed runtimes. The layer the data sites skip.
We synthesize; we don’t fabricate. Live pricing is attributed; benchmarks are sourced; vendor-reported figures are labelled as such and flagged pending independent reproduction.
OpenRouter
Live per-token pricing & availability (300+ models)
Artificial Analysis
Independent Intelligence Index, speed, latency
LMArena
Crowdsourced human-preference Elo
ARC Prize Foundation
ARC-AGI abstract reasoning benchmark
SWE-bench
Real-world software engineering tasks
Vendor model cards
Self-reported benchmarks (labelled as such)
There is no measured universal winner in this hub. The decision matrix proposes candidates for coding, games, writing, content and operational tasks. Select a model using the same task set, tools, context and acceptance criteria, then compare cost per accepted result. Our dedicated twelve-case Astra suite is defined but has not run.
OpenRouter supplies catalog pricing and routing; benchmark publishers provide their own evaluations. This hub connects selected model entries to workload candidates, creator workflows and public evaluation receipts. Editorial recommendations and third-party measurements retain their own evidence status.
Use the priced models in the explorer and calculator for a dated comparison. Gemini 3.5 Flash and Flash-Lite are different products. Token price alone cannot establish the cheapest accepted result: include failed attempts, tools and human correction.
Agent performance depends on the model together with its tools, memory, permissions and recovery strategy. Test completed work, tool errors, escalation, latency and total spending in that environment. Published benchmark scores can shortlist candidates; they do not establish a production winner for your workload.
Where a model maps to OpenRouter, pricing is fetched live (hourly) and marked with a ⚡ icon and "via OpenRouter." Otherwise it comes from our curated registry. Always verify against the provider before relying on it for billing.
Yes. Model rows with pricing provenance and registry evaluation status are available at /llm-hub.json. The versioned /llm-hub/manifest.json links models, editorial comparisons, candidate policies and public receipts. Treat editorial recommendations as hypotheses until task evidence qualifies them.