GPT-6 Astra Frontier escalation Proprietary Read briefing Highest-complexity reasoning, coding, research, tool use, and end-to-end professional work. 1.05M input · 128K output · material price boundary above 272K input $10/M input · $1/M cached · $50/M output. Batch/Flex 50% lower; long context and Fast mode carry premiums. Expect deliberate reasoning. Measure p95 with tools; reserve for high-value work where completion quality outweighs delay. Hosted APIProprietary Complex research Hard coding escalations Gated cybersecurity Cross-document judgement GPT-5.6 portfolio Three-tier routing ladder Proprietary Read briefing Luna for utility work, Terra for balanced generation, and Sol for complex professional tasks. 1.05M input · 128K output · long-context premium above 272K input Luna $0.20/$1.20, Terra $2/$12, Sol $4/$20 per M input/output. Sol pricing is promotional. Route for the product SLO: Luna for interactive volume, Terra for balanced work, Sol when additional reasoning is justified. Hosted APIProprietary Classification Extraction Drafting Tiered coding and analysis Claude Fable 5.1 Cache-efficient premium agent Proprietary Read briefing Demanding coding and knowledge work with unusually strong repeated-context cache economics and compaction. 1M input · 128K output; verify endpoint availability and regional terms $10/M input · $50/M output · $0.25/M cache reads. Cache writes break even after roughly 2–3 uses within TTL. No stable public figure in the briefing. Benchmark full agent runs, including tools and compaction, against your SLO. Hosted API / managed cloudProprietary Repository agents Repeated evidence packs High-value analysis Stable tool-schema workloads Gemini 3.8 Flash Cost-efficient agent workhorse Proprietary Read briefing Long-horizon coding and agents, multimodal input, built-in tools, tunable thinking, caching, Batch and priority tiers. 1,048,576 input · 65,536 output Introductory $0.75/M input and $3.75/M output through 31/12/2026; listed rates double on 01/01/2027. Designed for efficient work, but validate p50/p95 including thinking and tools. Use Batch only for delay-tolerant jobs. Hosted API / managed cloudProprietary Agentic workflows Multimodal context High-volume coding Batch enrichment gpt-oss 120B / 20B Permissive self-hosted reasoning Open source Read briefing Downloadable reasoning models for controlled environments, local adaptation, and infrastructure-owned inference. 128K reference context; validate serving engine and quantisation configuration No universal token price. Calculate accelerator, utilisation, throughput, platform, and operations cost. Hardware, precision, batching, concurrency, and serving engine determine latency; benchmark the deployed stack. Self-hosted or third-party inferenceApache 2.0 Private inference Controlled adaptation Edge or dedicated serving Apache-licensed deployments DeepSeek V4 family Large-context open-weight portfolio Open weight Read briefing Mixture-of-experts models for high-capability self-hosted or specialist-hosted inference. 1M reference context; validate model-card and serving configuration Compute-priced when self-hosted; third-party token rates vary by inference provider and region. Depends on active parameters, memory topology, quantisation, batching, and host capacity. Self-hosted or third-party inferenceDeepSeek model licence Sovereign deployment Large-context analysis Custom serving High-volume batch Llama 4 Scout / Maverick Broad-ecosystem open-weight models Open weight Read briefing Scout emphasizes very long context; Maverick provides a larger multimodal capability tier. Scout 10M · Maverick 1M reference context No universal token price. Hosted rates vary; self-hosted costs depend on the complete serving stack. Benchmark each host and quantisation. Headline context limits do not imply economical full-window operation. Self-hosted or third-party inferenceLlama 4 Community License Long-context retrieval Multimodal pipelines Cloud portability Custom evaluation Qwen and Mistral open families Efficient specialist deployment Open weight Read briefing Compact and mixture-of-experts options for multilingual, coding, edge, and throughput-sensitive workloads. Model-specific; verify exact model card, licence, and serving configuration No universal token price for downloaded weights. Compare measured accepted-output cost, not parameter count alone. Smaller active footprints can improve throughput, but quality gates and hardware fit determine the useful result. Self-hosted or third-party inferenceModel-specific; verify card Multilingual workloads Edge inference Coding assistants Cost-sensitive batch Frontier routing playbook Decision framework — not a model Framework Read briefing Five-route governance framework spanning utility, balanced, long-context, agentic, and frontier work. Not applicable. Enforces workload-specific context admission and budget thresholds. Normalizes input, cache, reasoning/output, tool, retry, fallback, regional, and priority costs across providers. Requires per-route p50/p95, retry, timeout, and failover evidence instead of relying on provider claims. Provider-neutralNot applicable Portfolio design Governance Evaluation planning Procurement and telemetry