Frontier Model Landscape

The proprietary frontier model families (Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok) as of November 2026: tiers, context windows, prices, and how vendors differentiate.

Contents

The frontier model landscape is the set of most capable proprietary LLMs you reach only through a vendor API or a cloud marketplace: Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini and xAI’s Grok. Each vendor ships a tiered family: a top reasoning model, a balanced workhorse, and one or two cheap, fast models. All of them share one API shape: token-metered pricing, a large context window, tool use, and an adjustable amount of reasoning. This page is a dated snapshot (read 2026-11-14). Names and prices change every few months, so check the vendor pages linked under Sources before you quote a number to a customer. For self-hostable alternatives, see open-weight models. For how to pick a model, see model selection.

The tier structure

Synthesis: every vendor now sells the same three-to-four-rung ladder, and the tier names map across vendors roughly as follows.

flowchart TB
  subgraph "Top: hardest reasoning, long-horizon agents"
    A1["Claude Fable 5.2"]
    A2["GPT-6 Astra"]
    A3["Gemini 4 Carbon"]
  end
  subgraph "Default workhorse: agentic coding, enterprise"
    B1["Claude Opus 5.5"]
    B2["GPT-6.1 Sol"]
    B3["Gemini 4 Argon"]
    B4["Grok 4.7"]
  end
  subgraph "Fast / everyday"
    C1["Claude Sonnet 5.5"]
    C2["GPT-6.1 Terra"]
    C3["Gemini 4 Hydrogen"]
  end
  subgraph "Cheapest: high-volume, routing, subagents"
    D1["Claude Haiku 5.5"]
    D2["GPT-6 Luna"]
    D3["Gemini 3.5 Flash-Lite"]
  end

Synthesis: this mapping is approximate. It is based on each vendor’s own positioning text and list price, not on head-to-head evals.

Snapshot table (list prices, USD per 1M tokens, read 2026-11-14)

VendorModel (API ID)ContextMax outputInput / OutputPositioning
AnthropicClaude Fable 5.2 (claude-fable-5-2)1M128K$8 / $40Demanding reasoning, long-horizon agentic work
AnthropicClaude Opus 5.5 (claude-opus-5-5)1M128K$4 / $20Recommended default; long-running agentic coding and knowledge work
AnthropicClaude Sonnet 5.5 (claude-sonnet-5-5)1M128K$2 / $10Speed and intelligence balance
AnthropicClaude Haiku 5.5 (claude-haiku-5-5)1M128K$0.10 / $0.50 (≤100K prompt)Classification, extraction, routing
OpenAIGPT-6 Astra (gpt-6-astra)1.05M128K$10 / $50Most capable
OpenAIGPT-6.1 Sol (gpt-6.1-sol)1.05M128K$2 / $10Near-Astra at lower cost
OpenAIGPT-6.1 Terra (gpt-6.1-terra)1.05M128K$1 / $5Visual reasoning, computer use, writing
OpenAIGPT-6 Luna (gpt-6-luna)1.05M128K$0.10 / $0.50High-volume, focused tasks
GoogleGemini 4 Carbon (gemini-4-carbon)1M64K (1M long decode)$2 / $10, rising to $4 / $20 on 2026-12-01Strongest reasoning, complex SWE
GoogleGemini 4 Argon (gemini-4-argon)1M64K (1M long decode)same as CarbonEfficient frontier reasoning
GoogleGemini 4 Hydrogen (gemini-4-hydrogen)1M64K (1M long decode)$1.50 / $7.50Everyday coding and agents
GoogleGemini 3.5 Flash-Liten/a on pagen/a$0.30 / $2.50Fastest, cheapest 3.5
xAIGrok 4.7 (grok-4.7)500Kn/a$2 / $6 (<200K prompt; $4 / $12 at or above 200K)Code and chat
xAIGrok 4.3 / 4.201Mn/a$1.25 / $2.50 (<200K)Reasoning, non-reasoning, multi-agent variants

For comparison, DeepSeek’s hosted API (serving open-weight checkpoints) charges $0.30 / $1.20 for deepseek-flash and $1.32 / $3.96 for deepseek-v4-pro at peak, with half price off-peak.

Knowledge cutoffs: June 2026 for all current Claude models; 2026-04-30 to 2026-06-01 for the GPT-6 family; May 2026 for Grok 4.7. The Gemini 4 Carbon page does not state one.

How the vendors differ

Reasoning control. All four expose a dial that trades thinking tokens for quality (see test-time compute and extended thinking):

  • Claude: adaptive thinking plus an effort parameter (low to max). On Fable 5.2 and Opus 5.5, thinking is always on and cannot be disabled. Default effort is medium on Opus 5.5 and Haiku 5.5, and high on Fable 5.2 and Sonnet 5.5.
  • OpenAI: reasoning effort from low to max on Astra and Sol. Terra and Luna also accept none. The o-series (o3, o3-pro) now appears only as legacy entries on the pricing page.
  • Gemini 4 Carbon: thinking levels low, medium and high.
  • xAI: separate reasoning and non-reasoning model IDs (grok-4.20).

Pricing mechanics that change the real bill:

LeverAnthropicOpenAIGooglexAI
Cached input0.1x base by default; 0.05x on Opus/Sonnet 5.5; 0.025x on Fable. Cache writes cost 1.25x (5 min) or 2x (1 h)0.1x on Astra and Luna; 0.05x on Sol and Terra0.05x on Gemini 4; 0.1x on Flash, Flash-Lite and 3.1 Pro; plus storage per hour on some models0.15x to 0.25x (e.g. $0.50 on Grok 4.7)
Batch50% offBatch/Flex 50% offBatch/Flex 50% offn/a on page
Premium speedFast mode, up to 2.5x output speed (Opus 5.5 at $8/$40, Fable 5.2 at $16/$80)Fast mode (renamed from Priority on 2026-07-30), 2xPriority about 1.8xn/a
Long-context stepNone on 4.6+ (except Haiku 5.5 above 100K)Above 272K input (e.g. Astra $20/$75)3.1 Pro above 200K≥200K prompt doubles the rate for all tokens
Data residencyinference_geo: "us" costs 1.1xRegional endpoints +10% (models released on or after 2026-03-05)n/an/a

See prompt caching, batch processing and cost and latency optimization.

Synthesis: list price per token is a poor proxy for cost. Cache-read discounts range from 4x (Grok 4.7) to 40x (Fable 5.2), and long-context steps can double a bill. Reasoning models also spend tokens on thinking (Google states that output prices include thinking tokens), so measure cost per completed task instead (see model selection).

Distribution. Claude is served on Anthropic’s API, Amazon Bedrock, Google Cloud (Google-managed and Anthropic-managed), Microsoft Foundry and Claude Platform on AWS, with platform-specific IDs (e.g. anthropic.claude-opus-5-5 on Bedrock). Partner clouds set their own prices and retirement dates. Gemini is on AI Studio and Vertex; see Amazon Bedrock, Google Vertex AI and Azure AI Foundry.

Modalities. All current Claude models take text and image input and return text. Gemini 4 adds audio, video and PDF input, and has separate Live, TTS, image and video models. GPT-6.1 Terra is positioned for visual reasoning and computer use. See multimodal models.

Access restrictions. Claude Mythos 5.2 has Fable 5.2’s capabilities but is open only to organisations verified through Anthropic programmes. Gemini 4 Carbon and Argon have no free tier, and free-tier data on other Gemini models may be used to improve Google products.

Lifecycle and versioning

  • Pinned IDs: every Claude model ID is a pinned snapshot, including the dateless IDs used from 4.6 on. xAI uses <model> (latest stable), <model>-latest and <model>-<date> (pinned). DeepSeek retires names and redirects them (e.g. deepseek-v4-flash is now served by V4.1-Flash-1017).
  • Retirement: Anthropic commits to retirement dates no sooner than September to November 2027 for the current models. Google will route Gemini 2.5 Pro and 2.5 Flash requests to 3.8 Flash and 3.5 Flash-Lite from 2026-12-01.
  • Breaking changes between generations are common: Opus 5.5, for example, breaks integrations that force tool use or turn thinking off. Always read the migration notes.
  • Scheduled price changes: Gemini 4 Carbon/Argon prices double on 2026-12-01, and Flash prices double on 2027-01-01. Budget for these.

Limitations and failure modes

  • Vendor benchmarks are not comparable. Each vendor picks its own harnesses and effort settings. Use benchmarks only as a shortlist filter, then run your own evals.
  • A large context window does not guarantee recall. A 1M-token window still suffers long-context degradation.
  • The snapshot decays fast. In 2026 each vendor shipped several generations. Synthesis: treat any model name in a design document as configuration, not architecture.

When to use which (practical)

Synthesis, grounded in each vendor’s own recommendation:

  • Default for agentic and coding work: Claude Opus 5.5 (Anthropic’s stated default), GPT-6.1 Sol, or Gemini 4 Argon. Pick by eval result and by ecosystem fit (which cloud the customer already uses, data residency requirements).
  • Hardest problems and multi-hour autonomy: Fable 5.2, GPT-6 Astra or Gemini 4 Carbon. Only escalate here after an effort sweep on the default model shows a gap.
  • High-volume extraction, classification and routing: Haiku 5.5, GPT-6 Luna or Gemini 3.5 Flash-Lite. All cost about $0.10 to $0.30 per million input tokens.
  • Native audio/video input or Google grounding: Gemini.
  • Self-hosting, data sovereignty or fine-tuning the weights: use open-weight models instead.

Sources

Suggest changes →