Claude Opus 5.5 from Anthropic, GPT-6 Astra from OpenAI and Grok 4.7 from xAI were all released in September 2026, and each provider publishes its limits, prices, access rules and safety material in a different shape. This page lines up the published facts for the three models with identical fields: API IDs, release dates, context and output limits, prices and long-prompt rules, tools, availability, restricted access, model and system cards, provider-reported benchmarks, lifecycle status and data handling. It is written for developers, architects, platform teams and buyers who need checked figures in one place before running their own evaluation. Providers are listed alphabetically (Anthropic, OpenAI, xAI) in every table; the order is not a ranking and the page gives no verdict. Every fact comes from the provider's own documentation, listed 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 releases within 19 days. GPT-6 Astra was released on 3 September 2026, Grok 4.7 on 21 September 2026 and Claude Opus 5.5 on 22 September 2026.
- Context windows differ by a factor of two. Anthropic lists 1M tokens with 128K output, OpenAI lists 1,050,000 tokens with 128,000 output, and xAI lists 500,000 tokens with no output limit stated.
- Each provider prices long prompts differently. Anthropic charges the standard rate across the full 1M window; OpenAI applies 2x input and 1.5x output above 272K input tokens; xAI doubles all rates once a prompt reaches 200,000 tokens.
- Batch discounts are not the same size. Anthropic and OpenAI list 50% off for batch work; xAI lists a 20% Batch API discount.
- No benchmark has a provider-reported score for all three. Each provider reports a different set of evaluations with its own settings, so this page lists them per provider rather than in a shared row.
How the models were selected
"GPT-6" is a family, not one model ID. OpenAI's GPT-6 generation has tiers named Astra, Sol, Luna and Terra, and its model page calls GPT-6 Astra (gpt-6-astra) "our most capable model for the most demanding work", so Astra represents GPT-6 here. The other two are single named models. Anthropic's models overview describes Claude Opus 5.5 as "for long-running agentic coding and knowledge work". xAI's model page describes Grok 4.7 as "SpaceXSI's frontier model for coding, agentic tasks, and knowledge work"; xAI's documentation now uses the name SpaceX Super Intelligence (SpaceXSI), while the API keeps the grok model IDs, and this page uses the xAI and Grok names that the API uses.
Each provider also publishes other models beside these. For tiers inside each family see Claude Opus 5.5 vs Opus 5 vs Opus 4.8, GPT-6 Astra vs Sol vs Luna vs Terra and Grok 4.7 vs Grok 4.6 vs Grok 4.5. For the same Anthropic and OpenAI models compared with Google instead of xAI see GPT-6 vs Claude Opus 5.5 vs Gemini 4 Carbon; for a wider field see Frontier AI Models Compared.
The three models at a glance
As of 2026-11-15. Sources: Anthropic models overview, OpenAI gpt-6-astra model page, xAI Grok 4.7 model page and release notes.
| Field | Claude Opus 5.5 (Anthropic) | GPT-6 Astra (OpenAI) | Grok 4.7 (xAI) |
|---|---|---|---|
| API ID | claude-opus-5-5 | gpt-6-astra | grok-4.7 |
| Status | Active | Current; only snapshot listed is gpt-6-astra | Current, on the public pricing page |
| Input modalities | Text, image | Text, image (audio and video not supported) | Text, image |
| Output modalities | Text | Text | Text |
| Knowledge cutoff | June 2026 (reliable knowledge cutoff) | 30 April 2026 | May 2026 (models page) |
| Reasoning control | Adaptive thinking, always on; effort default medium | reasoning.effort low, medium, high, xhigh, max | Reasoning effort low, medium, high (default), xhigh |
| Function calling and structured output | Tool use; structured outputs | Function calling; Structured Outputs | Function calling; structured outputs |
| Fine-tuning | Not stated by Anthropic on the pages checked | Not supported | Not stated by xAI on the pages checked |
| Open weights / licence | No / proprietary | No / proprietary | No / proprietary |
| Parameters | Not stated by Anthropic | Not stated by OpenAI | Not stated by xAI |
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. Model IDs follow different conventions (hyphens for Anthropic and OpenAI, a dot for xAI), which matters for configuration files and allowlists. See LLM Naming and Versioning Explained.
When was each model released?
| Model | Release | Snapshot naming | Source |
|---|---|---|---|
| Claude Opus 5.5 | 22 September 2026 | Undated ID; aliases not listed on the models overview | Anthropic announcement; models overview |
| GPT-6 Astra | 3 September 2026 (system card publication date) | Only gpt-6-astra is listed as a snapshot | OpenAI system card; model page |
| Grok 4.7 | 21 September 2026, on the xAI API and the US regional endpoint | No dated snapshot ID published; the models page says a plain name points to the latest stable version | xAI release notes; announcement |
xAI also lists Grok 4.7 Fast, which its release notes describe as the same model at twice the token rates (1.5x for long-context requests), offered only in Cursor and Grok Build and not on the public API. OpenAI and Anthropic each list a fast mode on their own APIs, covered in the pricing table. 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, OpenAI gpt-6-astra model page, xAI pricing page.
| Item | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| Input | $4 | $10.00 | $2.00 below 200K prompt tokens |
| Output | $20 | $50.00 | $6.00 below 200K prompt tokens |
| Cached input | $0.20 cache hit | $1.00 | $0.50 below 200K prompt tokens |
| Cache writes | $5 (5-minute); $8 (1-hour) | $12.50 | No separate write charge listed |
| Long prompts | No premium; full 1M window at standard rates | Over 272K input tokens: 2x input and cache rates, 1.5x output for the full request | At or above 200K prompt tokens: $4 input, $1 cached, $12 output for all tokens in the request |
| Batch | $2 / $10 (50% off) | 50% of Standard (Batch and Flex) | 20% discount on input, output, cached and reasoning tokens |
| Faster or priority tier | Fast mode $8 / $40 | Fast mode at 2x the applicable rates | Priority Processing at 2x standard, billed only when the response confirms the priority tier |
| US data location option | inference_geo "us" at 1.1x | Covered by a separate data residency guide | US regional endpoint at 1.1x global |
| Web search tool | $10 per 1,000 searches plus tokens | Supported; price not on the model page | $5 per 1,000 calls |
| Free tier | Not stated on the pricing page checked | Not supported | Not stated on the pricing page checked |
Worked example: illustrative arithmetic from the published standard rates, before reasoning-token differences, caching and tool calls.
| Request | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| 100,000 input and 10,000 output tokens | $0.60 | $1.50 | $0.26 |
| 300,000 input and 10,000 output tokens | $1.40 | $6.75 (long-prompt rates) | $1.32 (long-context rates) |
| 600,000 input and 10,000 output tokens | $2.60 | $12.75 (long-prompt rates) | Above the 500,000-token window |
The thresholds sit at different points (272K for OpenAI, 200K for xAI, none for Anthropic), so the relative cost of the three changes with prompt size. Each model also spends a different number of reasoning tokens per task, billed as output, so per-task cost needs your own measurement. OpenAI's model page also lists rate limits by usage tier (Free not supported; Build 5,000 requests and 1,000,000 tokens per minute), and xAI's model page lists 150 requests per second and 50,000,000 tokens per minute. See Model Pricing Explained, the LLM Pricing Tracker and LLM API Rate Limits Compared.
What is the context window of each model?
| Limit | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| Context window | 1M tokens | 1,050,000 tokens | 500,000 tokens |
| Max output | 128K tokens | 128,000 tokens | Not stated by xAI on the pages checked; release notes say no text output limit |
| Price step inside the window | None | 272K input tokens | 200,000 prompt tokens |
| Image input notes | Text and image input on all current models | Image input supported | Up to 20 MiB per image, JPG/JPEG and PNG (models page) |
Grok 4.7's window is half the size of the other two. Designs that summarise or retrieve to keep prompts below a price step change the cost profile for both GPT-6 Astra and Grok 4.7. See Context Windows Explained.
Capabilities and tools compared
| Capability | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| Web search and fetch | Web search and web fetch server tools | Web search in the Responses API | Web Search and X Search server-side tools |
| Code execution | Code execution tool (free with recent web search or fetch versions; otherwise 1,550 free hours per organisation per month) | Code interpreter and hosted shell | Code Execution tool |
| File and document search | Files API; search results | File search | File attachment search and Collections Search |
| MCP | MCP connector | MCP and tool search | Remote MCP (token-based pricing) |
| Computer use | Computer use tool | Computer use in the Responses API | Not stated by xAI on the pages checked |
| Reasoning output | Thinking cannot be turned off | Reasoning tokens billed as output | Responses API always returns reasoning.encrypted_content |
Server-side search, code and MCP tools act with the credentials and data the application gives them. See the Computer-Use and Browser Agents guide and Reasoning Models Explained.
Where is each model available?
| Channel | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| First-party API | Claude Platform (claude-opus-5-5) | OpenAI API | xAI API (us-east-1 listed; US regional endpoint us.api.x.ai) |
| Amazon Web Services | Amazon Web Services (announcement) | Amazon Bedrock (openai.gpt-6-astra) | Not named by xAI on the pages checked |
| Google Cloud | Google Cloud (announcement) | Private preview | Not named by xAI on the pages checked |
| Microsoft | Microsoft Azure (announcement) | Not stated on the model page checked | Not named by xAI on the pages checked |
| Other channels named | Claude subscriptions (Pro, Max, Team, Enterprise); Claude Code | Not stated on the model page checked | Cursor, Grok Build, third-party coding harnesses, model routers and cloud platforms (announcement, unnamed) |
Partner clouds set their own model IDs, regions and retirement schedules, so check each catalogue separately before relying on a row above.
Restricted-access variants and safeguards
- Anthropic: Opus 5.5 is generally available with no access request. The announcement states that most cybersecurity tasks are rerouted to Claude Opus 4.8, that biology safeguards match Claude Fable 5.1 with a Life Sciences Verification Program for vetted organisations, and that a preserved thinking anti-distillation safeguard applies to API accounts created on or after 31 August 2026.
- OpenAI: Astra is available in the API but not on the Free tier. Its system card states it is OpenAI's first model to reach the Critical level of cybersecurity capability under the Preparedness Framework, and lists Trusted Access for Cyber and Trusted Access for Biology Research sections.
- xAI: Grok 4.7 is on the public API. The announcement describes a new safeguard stack and states that select cybersecurity partners have invite-only access to Grok 4.7's red-team capabilities for defence research. The pricing page lists a $0.05 fee for requests blocked by the usage guidelines before generation on the Responses API.
Access reviews should record the exact model ID each project may call, since restricted programmes and general models share family names.
Model cards and system cards
| Item | Claude Opus 5.5 | GPT-6 Astra | Grok 4.7 |
|---|---|---|---|
| Document | Claude Opus 5.5 System Card (referenced in the announcement) | GPT-6 Astra system card, published 3 September 2026, updated to 30 October 2026 | Not stated by xAI: no model card or system card is linked from the model page or announcement checked |
| Framework cited | Not restated on the pages checked for this comparison | Preparedness Framework | Not stated by xAI on the pages checked |
| Headline safety statements | Anthropic states the best score of any model it tested on its automated behavioural audit; watermarking for EU AI Act compliance | Critical cybersecurity capability; Gray Swan IPI Arena attack success 8.5% | LatchBio biosafety benchmark 62.4%; HackerBench v0.3: 3.3% of risky dual-use prompts allowed through (announcement) |
Each provider uses its own terms and tests, so these statements 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. Terminal-Bench 4.0 appears in both Anthropic's and xAI's announcements, but OpenAI's Astra system card does not report it, and the two runs use different effort settings. The table lists three figures each provider reports for its own model.
| Provider and model | Evaluation and score as reported | Source and date |
|---|---|---|
| Anthropic: Claude Opus 5.5 | Terminal-Bench 4.0: 66.4% (xhigh effort); CursorBench 4.0: 57.8%; GDPval-AA v2.1: 1846 Elo | Anthropic announcement, 22 Sep 2026 |
| OpenAI: GPT-6 Astra | HealthBench Professional: 64.7 (length-adjusted); HealthBench Hard: 36.6; Gray Swan IPI Arena attack success: 8.5% (lower is better) | GPT-6 Astra system card, 3 Sep 2026 |
| xAI: Grok 4.7 | Terminal-Bench 4.0: 37.6% (xhigh); CursorBench 4.0: 46.3% (xhigh); HealthBench Professional: 56.7% (xhigh) | xAI announcement, 21 Sep 2026 |
xAI's announcement also shows scores for competitor models, including a GDPval Elo for GPT-6 Astra; those competitor figures are xAI-reported and are not reproduced here. Similarly named evaluations can differ in version and harness (for example "GDPval-AA v2.1" in Anthropic's table and "GDPval Elo" in xAI's). Independent leaderboard results (LMArena, Artificial Analysis, Epoch AI) were not checked for this page and are not included. See How to Read LLM Benchmarks.
Deprecations and retirements
| Model | Details | Source |
|---|---|---|
| claude-opus-5-5 | Active; tentative retirement not sooner than 22 September 2027; at least 60 days' notice for publicly released models | Anthropic model deprecations page |
| gpt-6-astra | Not in any deprecation table (2026-11-15); generally available models get at least 6 months' notice unless safety or compliance requires faster | OpenAI deprecations page |
| grok-4.7 | No retirement announced on the pages checked (2026-11-15); 2026 retirement entries on the release notes concern voice and image models | xAI release notes |
Track all three on the LLM Deprecation and Retirement Tracker.
Data handling and security
| Topic | Anthropic | OpenAI | xAI |
|---|---|---|---|
| Training on API data | Retained data "is never used for model training without your express permission" | API data not used to train or improve models unless the customer opts in | Does not train on API inputs or outputs without explicit permission |
| Default retention | Batch 29 days; code execution containers up to 30 days; flagged content up to 2 years | Abuse monitoring logs up to 30 days; Responses API state at least 30 days when stored | Requests and responses encrypted at rest, kept 30 days for abuse auditing, then deleted |
| Zero data retention | Per organisation through Anthropic sales; batch, Files API and code execution not eligible | ZDR or Modified Abuse Monitoring with prior OpenAI approval; Batches and Files not eligible | Team-wide toggle in the Console; disables stateful Responses API, Files, Collections, Batch API and deferred completions |
| Data location | inference_geo "us" at 1.1x | Regional storage and processing options in a separate guide | us.api.x.ai keeps processing and retained request data in the US at 1.1x; serves grok-4.7 and grok-4.6 only |
| Assurance stated on the page checked | Not stated on the ZDR page | Not stated on the data controls page | SOC 2 Type 2; BAA by questionnaire |
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 OpenAI API Keys Security Guide, LLM Data Retention and Training Policies Compared and Zero Data Retention for AI.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Prompts above 500,000 tokens | Opus 5.5 and Astra list windows of about one million tokens; Grok 4.7 lists 500,000 |
| Prompt size and price steps | None for Opus 5.5; 272K for Astra; 200K for Grok 4.7 |
| Batch-heavy workloads | 50% off for Opus 5.5 and Astra; 20% off for Grok 4.7; Batch is outside ZDR for all three |
| Published output limit | 128K (Opus 5.5) and 128,000 (Astra); not stated for Grok 4.7 |
| Cloud already in use | Check the partner cloud rows above and that cloud's own catalogue |
| Safety documentation required by policy | System cards for Opus 5.5 and Astra; none linked for Grok 4.7 on the pages checked |
| Lifecycle dates | Opus 5.5: not sooner than 22 September 2027. Astra and Grok 4.7: no retirement date stated |
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, gpt-6-astra, grok-4.7) each application calls in the AI inventory; see the AI Register and AI Inventory Template.
- Map typical prompt sizes against the 200K and 272K price steps before forecasting spend.
- Measure reasoning tokens per task at the effort level you plan to run, since all three bill them as output.
- Confirm whether batch, files and stateful features you rely on fall outside each provider's zero data retention arrangement.
- Check partner cloud and router catalogues for their own model IDs, regions and retirement dates.
- Store every provider API key in a secrets manager, scope it to one project and set spend alerts; see How to Set Spend Limits on LLM API Keys.
- Re-check this page's figures against the provider pages on the day you commit budget.
Standards and references
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Claude Opus 5.5 announcement (accessed 2026-11-15)
- Anthropic: Model deprecations (accessed 2026-11-15)
- Anthropic: Zero data retention (accessed 2026-11-15)
- OpenAI: GPT-6 Astra model page (accessed 2026-11-15)
- OpenAI: GPT-6 Astra system card (accessed 2026-11-15)
- OpenAI: Deprecations (accessed 2026-11-15)
- OpenAI: Data controls in the OpenAI platform (accessed 2026-11-15)
- xAI: Grok 4.7 model page (accessed 2026-11-15)
- xAI: Pricing (accessed 2026-11-15)
- xAI: Release notes (accessed 2026-11-15)
- xAI: Grok 4.7 announcement (accessed 2026-11-15)
- xAI: API security FAQ (accessed 2026-11-15)
Related NHI Mgmt Group resources: Claude Opus 5.5, GPT-6 Astra, Grok 4.7, xAI Grok Models, Frontier AI Models Compared.