Claude Opus 5.5 from Anthropic, Gemini 4 Carbon from Google and GPT-6 Astra, the top tier of OpenAI's GPT-6 generation, were all released between September and November 2026, and each comes with its own limits, prices, access rules and safety documentation. This page puts the published facts for the three models side by side with identical fields: API IDs, release dates, context and output limits, prices, tools, availability, restricted variants, model and system cards, provider-reported benchmarks, lifecycle status and data handling. It is written for developers, architects and buyers who need checked figures in one place. Providers are listed alphabetically (Anthropic, Google, OpenAI) in every table; the order is not a ranking, and the page gives no verdict. Every fact comes from the provider's own documentation, cited under Standards and references.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model details change, so confirm against each provider's current documentation.
Key takeaways
- Three models, three release dates. Claude Opus 5.5 launched on 22 September 2026, Gemini 4 Carbon became generally available on 10 November 2026 and GPT-6 Astra was released on 3 September 2026.
- All three have a context window of about one million tokens. Anthropic lists 1M tokens with 128K output; Google lists 1,048,576 tokens shared by input and output; OpenAI lists 1,050,000 tokens with 128,000 output.
- List prices differ by provider and by date. Per million tokens: Opus 5.5 $4 input and $20 output; Carbon $2 and $10 until 30 November 2026, then $4 and $20; Astra $10 and $50, with higher rates for prompts over 272K tokens.
- Each provider runs a restricted cyber programme. Anthropic offers a Cyber Verification Program, Google an allowlisted gemini-4-carbon-cyber-permissive variant through its Fairwind Program, and OpenAI separate cyber models under trusted access.
- No benchmark has a provider-reported score for all three. Each provider reports its own evaluations, with different harnesses and settings, so this page lists them per provider rather than in a shared row.
How the models were selected
"GPT-6" is not a single model ID. OpenAI's latest-model guide describes "the GPT-6 model family" with tiers named Astra, Sol, Terra and Luna, and says: "Use GPT-6 Astra for the most demanding work." OpenAI's model page calls Astra "our most capable model for the most demanding work", so this page uses GPT-6 Astra (gpt-6-astra) to represent GPT-6. The other two are named models: Anthropic's models overview describes Claude Opus 5.5 as "for long-running agentic coding and knowledge work", and Google's Carbon announcement calls Carbon "our strongest model with frontier performance in cybersecurity capabilities".
Each provider also publishes models positioned above or beside these. Anthropic suggests Claude Fable 5.2 for demanding reasoning and long-horizon agentic work. Google states that "Argon remains the ideal choice for the majority of software engineering and knowledge work". OpenAI positions GPT-6.1 Sol for complex work at lower cost. For tiers inside each family see Claude Opus 5.5 vs Opus 5 vs Opus 4.8, Gemini 4 Carbon vs Argon vs Hydrogen and GPT-6 Astra vs Sol vs Luna vs Terra; for a wider field of providers see Frontier AI Models Compared.
The three models at a glance
As of 2026-11-15. Sources: Anthropic models overview, Google Cloud Gemini 4 Carbon model page, OpenAI gpt-6-astra model page.
| Field | Claude Opus 5.5 (Anthropic) | Gemini 4 Carbon (Google) | GPT-6 Astra (OpenAI) |
|---|---|---|---|
| API ID | claude-opus-5-5 | gemini-4-carbon | gpt-6-astra |
| Status | Active | GA | Flagship, listed first in OpenAI's models page |
| Input modalities | Text, image | Text, image, audio, video | Text, image |
| Output modalities | Text | Text | Text |
| Knowledge cutoff | June 2026 (reliable knowledge cutoff) | March 2026 | 30 April 2026 |
| Reasoning control | Adaptive thinking, always on; effort parameter, default medium | thinking_level LOW, MEDIUM (default), HIGH | Reasoning effort low, medium, high, xhigh, max |
| Sampling parameters | Not stated on the models overview | temperature, top_p, top_k fixed; custom values ignored | temperature, top_p and top_logprobs must be removed when reasoning effort is not none (latest-model guide) |
| Open weights / licence | No / proprietary | No / proprietary | No / proprietary |
| Parameters | Not stated by Anthropic | Not stated by Google | Not stated by OpenAI |
None of the three providers publishes open weights or a parameter count for these models, so "Is any of them open source?" has the same answer for all three: no. See LLM Naming and Versioning Explained for how each provider names snapshots.
When was each model released?
| Model | Release | Snapshot naming | Source |
|---|---|---|---|
| Claude Opus 5.5 | 22 September 2026, as the first model in the Claude 5.5 family | Undated ID; Anthropic states such an ID is the snapshot, not an alias | Anthropic release notes and announcement |
| Gemini 4 Carbon | 10 November 2026 (GA); a cyber variant was offered to defenders through the Fairwind Program before GA | No dated snapshot suffix published | Google Cloud model page; Carbon announcement |
| GPT-6 Astra | 3 September 2026 | Only gpt-6-astra is listed as a snapshot | OpenAI API changelog; model page |
For a dated list across all providers see the LLM Release Tracker.
How much does each model cost?
As of 2026-11-15. USD per million tokens, standard first-party API rates. Sources: Anthropic pricing page, Google AI Studio pricing page, OpenAI gpt-6-astra model page.
| Item | Claude Opus 5.5 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|
| Input | $4 | $2.00 through 30 Nov 2026; $4.00 from 1 Dec 2026 | $10.00 |
| Output | $20 | $10.00; then $20.00 (includes thinking) | $50.00 |
| Cached input | $0.20 cache hit | $0.10; then $0.20 (implicit caching only) | $1.00 |
| Cache writes | $5 (5-minute); $8 (1-hour) | No separate write charge listed | $12.50 (1.25x input) |
| Batch | $2 / $10 (50% discount) | $1.00 / $5.00; then $2.00 / $10.00 (Flex the same) | 50% of Standard (Batch and Flex) |
| Faster or priority tier | Fast mode $8 / $40 | Priority $3.60 / $18.00; then $7.20 / $36.00 | Fast mode at 2x the applicable rates |
| Long prompts | No premium; full 1M window at standard rates | No long-context tier listed | Over 272K input tokens: 2x input and cache rates, 1.5x output for the full request |
| Free tier | Not stated on the pricing page checked | Not available on the Free tier | Free tier not supported |
Availability notes: Anthropic applies a 1.1x multiplier when inference_geo is set to "us"; Google bills every processed token, including thinking tokens and grounding tool instructions; OpenAI lists rate limits for its Build, Launch and Grow tiers. Illustrative arithmetic for 100,000 input and 10,000 output tokens at standard rates, before reasoning-token differences: Opus 5.5 $0.60; Carbon $0.30 until 30 November 2026 and $0.60 from 1 December 2026; Astra $1.50. Each model spends a different number of reasoning tokens per task, so per-task cost needs your own measurement. See Model Pricing Explained and the LLM Pricing Tracker.
What is the context window of each model?
| Limit | Claude Opus 5.5 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|
| Context window | 1M tokens | 1,048,576 tokens, shared by input and output | 1,050,000 tokens |
| Max output | 128K tokens | 65,536 per streamed segment, 32,768 non-streamed; up to 1,048,576 minus input with long decode | 128,000 tokens |
| How to check programmatically | Models API returns live max_input_tokens | Count Tokens API | Model page lists limits per snapshot |
Window size is similar across the three; the differences are in output limits and in how long prompts are priced. See Context Windows Explained.
Capabilities and tools compared
| Capability | Claude Opus 5.5 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|
| Function calling | Tool use with tool_choice auto and strict tool use | Function calling | Function calling |
| Structured output | Strict tool use | Structured output | Structured Outputs |
| Computer use | computer_toolset_20260801 on the Claude API and Google Cloud | Supported on gemini-4-carbon | Computer use tool in the Responses API |
| Web search and code execution | Web search, web fetch and code execution server tools | Grounding with Google Search and Maps, code execution, URL context | Web search, code interpreter, hosted shell |
| Other documented tools | Bash and text editor tools | Agentic video understanding, system instructions | File search, image generation, apply patch, skills, MCP, tool search |
| APIs | Messages API | GenerateContent, Interactions, Count Tokens, Chat Completions | Responses, Chat Completions, Batch and others |
Computer use and hosted shell give a model reach into systems; each connected tool acts with credentials. See the Computer-Use and Browser Agents guide.
Where is each model available?
| Channel | Claude Opus 5.5 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|
| First-party API | Claude API | Gemini API paid tier (Google AI Studio, Android Studio) | OpenAI API |
| Amazon Web Services | Claude in Amazon Bedrock; Claude Platform on AWS | Not stated by Google on the pages checked | Amazon Bedrock (openai.gpt-6-astra) |
| Google Cloud | Claude Models on Google Cloud | Gemini Enterprise Agent Platform, verified enterprise customers; global, us, eu | Private preview |
| Microsoft | Claude in Microsoft Foundry | Not stated by Google on the pages checked | Not stated on the model page checked |
| Consumer and business apps | Claude apps (Pro, Max, Team, Enterprise) | Gemini app, AI Mode and Antigravity for Google AI Ultra; Gemini for business | Not stated on the model page checked |
Partner clouds set their own model IDs and retirement schedules, so check each catalogue separately.
Restricted-access variants and safeguards
- Anthropic: Opus 5.5 is generally available with no access request. Its system card describes cyber classifiers in which blocked requests fall back to Opus 4.8, and a Cyber Verification Program that offers reduced restrictions to verified users.
- Google: gemini-4-carbon-cyber-permissive is allowlist-only through the Fairwind Program, defaults to HIGH thinking, does not support computer use and runs in the global and us regions; access to earlier Gemini cyber models does not carry over.
- OpenAI: Astra is not restricted in the API but is not on the Free tier. OpenAI's restricted cyber models, such as gpt-6-cyber, sit separately under trusted access, and the Astra system card describes a more conservative refusal boundary for users flagged as potentially high risk.
Access reviews should record the exact model ID each project may call, since restricted and general variants share family names.
Model cards and system cards
| Item | Claude Opus 5.5 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|
| Document | System Card: Claude Opus 5.5 (22 September 2026) | Gemini 4 Carbon model card and Gemini 4 Carbon Safety and Security report | GPT-6 Astra system card (3 September 2026) |
| Framework cited | Responsible Scaling Policy and Frontier Compliance Framework | Frontier Safety Framework (in the safety report) | Preparedness Framework |
| Headline statement | Chemical and biological risk treated as CB-1 but not CB-2 | Strengthened safeguards for CBRN and cyber misuse; prompt injection robustness "remains an area for continued improvement" | "Our first model to reach the Critical level of cybersecurity capability"; treated as High in the biological and chemical domain |
Each framework uses its own levels, so the ratings cannot be lined up against each other. See Model Cards and System Cards Explained and the Frontier AI Model Risk Frameworks Guide.
Benchmarks
No evaluation on the pages checked has a score reported by each provider for its own model, so there is no shared row. The table lists three figures each provider reports for its own model, from its own document. Similar names hide different settings: Anthropic's "OSWorld 2.1 (partial)" and Google's "OSWorld-2.1 (offline subset)" are not the same run, so the figures are not comparable.
| Provider and model | Evaluation and score as reported | Source and date |
|---|---|---|
| Anthropic: Claude Opus 5.5 | Terminal-Bench 4.0: 66.4%; OSWorld 2.1 (partial): 81.8%; Humanity's Last Exam (with tools): 67.7% | Anthropic announcement, 22 Sep 2026 (max effort) |
| Google: Gemini 4 Carbon | Terminal-Bench 4.0 (AA): 65.2%; OSWorld-2.1 (offline subset): 81.5%; SEC-bench Pro: 88.2% | Google DeepMind Gemini page, undated, as shown 2026-11-15 |
| OpenAI: GPT-6 Astra | SEC-Bench Pro pass@1: 85.4%; Gray Swan IPI Arena attack success: 8.5% (lower is better); HealthBench Professional: 64.7 | GPT-6.1 Sol system card addendum, 29 Sep 2026; GPT-6 Astra system card |
Google's DeepMind page also shows its own table that includes scores for Claude Opus 5.5 and GPT-6 Astra; those competitor figures are Google-reported and are not reproduced here. Independent leaderboard results are not included. See How to Read LLM Benchmarks.
Deprecations and retirements
| Model | Details | Source |
|---|---|---|
| claude-opus-5-5 | Active; retirement not sooner than 22 September 2027; at least 60 days' notice for publicly released models | Anthropic model deprecations page |
| gemini-4-carbon | GA; retirement date not stated by Google on the pages checked (2026-11-15); price rise on 1 December 2026 | Google Cloud model page; AI Studio pricing |
| gpt-6-astra | No notice date, shutdown date or replacement listed (2026-11-15) | OpenAI deprecations page |
Track all three on the LLM Deprecation and Retirement Tracker.
Data handling and security
| Topic | Anthropic | OpenAI | |
|---|---|---|---|
| Training on API data | Retained data "is never used for model training without your express permission" | Paid-tier Standard, Batch, Flex and Priority content not used to improve products; Contributor tier content is | API data not used to train or improve models unless the customer opts in |
| Zero data retention | Per organisation through Anthropic sales; covers eligible Messages and Token Counting features | Data residency, CMEK, VPC Service Controls and Access Transparency on Google Cloud | Zero Data Retention and Modified Abuse Monitoring with prior OpenAI approval |
| Default retention notes | Batch data 29 days; code execution containers up to 30 days; flagged content up to 2 years | Implicit caching on by default for shared prefixes of 4,096 tokens or more | Abuse monitoring logs up to 30 days; Responses API state 30 days by default |
| Data location | inference_geo "us" at 1.1x | global, us and eu; ML processing in us and eu | Separate data residency guide |
Keys and service accounts that call these models are non-human identities with spend and data exposure. See the Anthropic API Keys Security Guide, the Google Gemini API and AI Studio Security Guide, the OpenAI API Keys Security Guide and Zero Data Retention for AI.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Input types needed | Audio and video input: Carbon. Text and image: all three |
| Output length | 128K (Opus 5.5), 128,000 (Astra), up to the full window with long decode (Carbon) |
| Cost through and after 30 November 2026 | Carbon's list price doubles on 1 December 2026; Opus 5.5 and Astra list no scheduled change |
| Prompts above 272K tokens | Astra applies higher rates; Opus 5.5 and Carbon list none |
| Cloud already in use | Check the partner cloud rows above and that cloud's own catalogue |
| Data retention requirement | Compare each provider's retention and approval steps in the table above against your policy |
| Lifecycle dates | Opus 5.5: not sooner than 22 September 2027. Carbon and Astra: no retirement date stated on the pages checked |
Run your own evaluation on representative tasks with each model's recommended reasoning setting before deciding.
Practitioner checklist
- Record which exact model ID (claude-opus-5-5, gemini-4-carbon, gpt-6-astra) each application calls, in the AI inventory.
- Remove sampling parameters that a model ignores or rejects before switching providers, and test for HTTP 400 errors.
- Re-forecast Carbon spend for 1 December 2026 and Astra spend for prompts over 272K tokens.
- Measure reasoning tokens per task at the effort or thinking level you plan to run, since they are billed as output.
- Confirm the retention terms and any zero data retention approval with each provider before sending sensitive data.
- Check partner cloud catalogues for their own model IDs and retirement dates.
- Store every provider API key in a secrets manager, scope it to one project and set spend alerts.
Standards and references
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Zero data retention (accessed 2026-11-15)
- Google Cloud: Gemini 4 Carbon model page (accessed 2026-11-15)
- Google AI Studio: Gemini API pricing (accessed 2026-11-15)
- Google: Gemini 4 Carbon is now available (accessed 2026-11-15)
- Google DeepMind: Gemini models (accessed 2026-11-15)
- OpenAI: GPT-6 Astra model page (accessed 2026-11-15)
- OpenAI: Using GPT-6 and GPT-6.1 (accessed 2026-11-15)
- OpenAI: Data controls in the OpenAI platform (accessed 2026-11-15)
Related NHI Mgmt Group resources: Claude Opus 5.5, Gemini 4 Carbon, GPT-6 Astra, Frontier AI Models Compared, LLM Release Tracker.