GPT-6 Astra
For complex multi-step engineering, research, professional work, computer use and high-impact assignments where more capability can justify a higher rate.
- Model ID
gpt-6-astra- Input / 1M
- $10
- Output / 1M
- $50
- Reasoning
- low–max
MODEL INTELLIGENCE
GPT-6 Models
Astra vs Sol vs Luna.
Compare GPT-6 Astra, GPT-6.1 Sol and Luna: verified pricing, 1.05M-token context, working token-cost examples, reasoning options and API guidance.
21tracked signals
1primary sources
November 5, 2026latest tracked
MODEL FAMILY REFERENCE · VERIFIED SOURCES
What is GPT-6? Understand the family before choosing a model.
GPT-6 is OpenAI's family of reasoning-focused models for coding, professional knowledge work, research, and tool-using agents. The main model-selection path currently has three tiers: GPT-6 Astra for the hardest tasks, GPT-6.1 Sol for near-Astra capability at a lower price, and GPT-6 Luna for focused, high-volume workloads.
They share a 1,050,000-token context window, a 128,000-token maximum output, text and image input, and text output. Shared limits do not mean interchangeable behavior: reasoning controls, cache economics, quality under realistic workloads and API constraints are different. Use this page to compare published facts and choose an initial candidate; use the linked individual model pages for detailed endpoint and release information.
Important 2026 update: OpenAI announced GPT-6.1 Sol on September 29 as the newer middle-tier model. The earlier GPT-6 Sol still has a distinct model ID and reference page. They are not synonyms. This guide compares the three models highlighted in OpenAI's current GPT-6 selection guidance.
Editorial source check: October 13, 2026 · Standard API list prices in USD · All rates and model constraints sourced from OpenAI developer documentation.
THREE CURRENT MODEL CHOICES
GPT-6 Astra vs GPT-6.1 Sol vs GPT-6 Luna.
Choose by required quality, expected task volume and cost per successful result—not just model branding.
GPT-6.1 Sol
Near-Astra positioning for complex coding agents, workflows and computer use at one-fifth Astra's standard input and output token rates.
- Model ID
gpt-6.1-sol- Input / 1M
- $2
- Output / 1M
- $10
- Reasoning
- low–max
GPT-6 Luna
For structured extraction, classification, repeated support tasks and cost-sensitive automation when your own evaluation confirms adequate quality.
- Model ID
gpt-6-luna- Input / 1M
- $0.1
- Output / 1M
- $0.5
- Reasoning
- none–max
Why not GPT-6 Sol (original) in the primary tier cards? OpenAI's newest recommendation guide now selects GPT-6.1 Sol. The original GPT-6 Sol remains relevant for migrations, cost comparisons, compatibility and previously tested workloads.
OFFICIAL STANDARD API PRICING
GPT-6 API pricing: input, cached input, cache writes and output.
Per million tokens, before any extra tool, data-residency or specialized serving charges. Cached reads and cache writes are distinct billing dimensions.
Short-context rates: up to 272,000 input tokens
| Model | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| GPT-6 Astra | $10 | $1 | $12.5 | $50 |
| GPT-6.1 Sol | $2 | $0.1 | $2.5 | $10 |
| GPT-6 Luna | $0.1 | $0.01 | $0.125 | $0.5 |
Long-context rates: over 272,000 input tokens
| Model | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| GPT-6 Astra | $20 | $2 | $25 | $75 |
| GPT-6.1 Sol | $4 | $0.2 | $5 | $15 |
| GPT-6 Luna | $0.2 | $0.02 | $0.25 | $0.75 |
How the threshold works: when total prompt input exceeds 272K tokens, the entire request uses the higher long-context price. The published schedule is 2× input and cache rates, and 1.5× output rates. A 300K-token prompt does not pay standard rates on the first 272K and a higher rate on only the remaining 28K.
Rates can also change by service tier: Batch, Flex, Fast and Ultrafast may have their own availability and prices. Data-residency processing may add a regional uplift, and some tool calls are billed separately. Always consult OpenAI's official rate card for your deployment. The latest Sol model has $0.1/M cached input; original GPT-6 Sol has $0.2/M at short context.
WORKLOAD COST ESTIMATOR
Calculate a GPT-6 API request cost.
Change the billable token counts to estimate the same workload across Astra, 6.1 Sol and Luna. This calculator deliberately uses Standard token rates only.
Standard short-context rates · 8,000 prompt tokens
The estimator excludes tool calls, taxes, Fast/Flex/Batch/Ultrafast tiers, extra compute, residency uplifts and retry cost. Models may use different output-token counts for the same task. Values above the documented 1,050,000-token input or 128,000-token output limits are rejected.
Worked example prices for three different token workloads
| Example request | Astra | 6.1 Sol | Luna |
|---|---|---|---|
| Short Q&A8K new input, 2K output | $0.18 | $0.036 | $0.0018 |
| Cached agent100K new, 40K cached, 10K output | $1.54 | $0.304 | $0.0154 |
| Long document300K new input, 20K output | $7.5 | $1.5 | $0.075 |
These calculations apply OpenAI's published Standard rates to fixed hypothetical token counts; they are not claims about actual task quality, completion time or token usage across models.
PRODUCTION MODEL SELECTION
Which GPT-6 model is best for your use case?
The most useful choice depends on how much work succeeds, the cost of an error, and how often requests repeat.
Benchmark 6.1 Sol, then Astra
Use GPT-6.1 Sol as a quality/cost baseline for multi-file edits, tests, tool orchestration and debugging. Escalate genuinely hard cases to Astra only when your evaluation shows better outcomes.
Consider Luna with validation
For categorization, structured extraction and routine internal automation, Luna can reduce spend. Verify schema conformance, numeric accuracy and failure handling before routing large traffic volumes.
Pay for demonstrated quality
For sensitive documents, expert research or difficult analytical synthesis, compare the full results on source accuracy and downstream rework, not only the length of the generated answer.
Shared window, different judgment
All three current models offer 1.05M context. Test retrieval and reasoning across long prompts; avoid assuming a large input limit means guaranteed factual recall, and account for the long-context price band.
Use appropriate API permissions
Agents need tool controls, limited permissions, monitoring and human approval for irreversible actions. OpenAI's Responses API is the recommended surface for reasoning with tools.
Choose processing separately
Fast and other processing options vary by model, region and account. Lower reasoning effort may reduce latency, but performance and residency eligibility must be measured against your own constraints.
1
Build a representative test set including difficult and failure-prone tasks.
2
Establish quality and cost per successful job using 6.1 Sol.
3
Use Luna where measured success remains high at scale.
4
Route hard or high-impact tasks to Astra when it earns the premium.
API REFERENCE SPECS
GPT-6 token limits, knowledge cutoff and reasoning modes.
| Specification | Astra | 6.1 Sol | Luna |
|---|---|---|---|
| Model ID | gpt-6-astra | gpt-6.1-sol | gpt-6-luna |
| Context window | 1,050,000 | 1,050,000 | 1,050,000 |
| Max output | 128,000 | 128,000 | 128,000 |
| Knowledge cutoff | 2026-04-30 | 2026-04-30 | 2026-05-18 |
| Reasoning effort | low, medium, high, xhigh, max | low, medium, high, xhigh, max | none, low, medium, high, xhigh, max |
| Modalities | Text + image in / text out | Text + image in / text out | Text + image in / text out |
Reasoning isn't a single switch. Astra and 6.1 Sol do not support reasoning.effort=none, while the earlier Sol and Luna do. All three current models support medium-effort starting configurations; high, xhigh and max are valuable only if your own task tests justify the extra tokens or latency. OpenAI documents tool use and streaming, but support depends on the particular API endpoint and parameter combination.
PYTHON DEVELOPER EXAMPLE
How to call GPT-6.1 Sol with OpenAI's Responses API.
The example explicitly names a current model. Switch model to gpt-6-astra or gpt-6-luna only after checking applicable reasoning effort and endpoint rules.
from openai import OpenAI
client = OpenAI() # OPENAI_API_KEY in environment
response = client.responses.create(
model="gpt-6.1-sol",
reasoning={"effort": "medium"},
input="Review this function for correctness and edge cases.",
)
print(response.output_text)Preferred for reasoning with tools
Use Responses for tool-enabled workflows. Astra and 6.1 Sol require it for tool calling; the older Sol and Luna only support Chat Completions function calling with reasoning set to none.
Measure hits and writes
Separate input, cached input and cache-write tokens. Reuse stable prefixes where possible; inspect actual usage rather than assuming caching automatically improves every request.
Remove unsupported options
With reasoning enabled, parameters such as temperature and top_p may be invalid. Check OpenAI's current migration guidance before copying older SDK code.
For pricing and architecture, account for the difference between OpenAI API model selection and the GPT-6 models exposed through ChatGPT, Codex or other products. Product availability and API billing are separate.
REAL-WORLD RESTRICTIONS
GPT-6 limitations and deployment checklist.
01
A million tokens is not a guarantee of perfect recall. Test extraction and answer faithfulness with long, messy documents and irrelevant distractors.
02
The cheapest token is not always the cheapest result. Include retries, tool calls, reasoning token use, caching and human review.
03
Vendor evaluations are not deployment guarantees. Look for benchmark methodology and match the prompt, effort, tools and grading to your real workloads.
04
Models and APIs differ. Do not assume a Chat Completions call with tools or the same sampling parameters works with all tiers.
05
Long context changes the entire request bill. Crossing 272K prompt tokens can materially change cost; use a high-context test budget.
06
Agents need security boundaries. Use permission scoping, audit logs, prompt-injection defenses and approvals for sensitive or irreversible actions.
FAMILY RELEASE TRACKER
GPT-6 release timeline and the newer Sol model.
Astra: OpenAI launched its quality-first GPT-6 model for difficult professional and technical work. Official announcement ↗
Sol and Luna: new pricing and efficiency choices joined the GPT-6 family. Official announcement ↗
6.1 Sol: the middle tier was upgraded. Short-context cached input fell from $0.2 per million on original Sol to $0.1 on 6.1 Sol. Official upgrade ↗
For existing integrations, benchmark gpt-6.1-sol rather than silently replacing gpt-6-sol. Recheck tool calls, reasoning effort, cache economics, latency and evaluation outcomes. SXF keeps individual original Sol and new Sol pages separately.
COMMON QUESTIONS
GPT-6 FAQ: models, costs and API compatibility.
What is GPT-6?
GPT-6 is OpenAI's family of large-context reasoning models. Its current main choices are GPT-6 Astra for the hardest tasks, GPT-6.1 Sol for near-Astra work at lower cost, and GPT-6 Luna for efficient high-volume requests.
Is GPT-6 one model or a family?
It is a family, not a single interchangeable API model ID. Use gpt-6-astra, gpt-6.1-sol, or gpt-6-luna; the earlier gpt-6-sol is separately documented.
What is the latest GPT-6 Sol model?
GPT-6.1 Sol was introduced on September 29, 2026 as the updated middle-tier choice. Original GPT-6 Sol remains documented separately and may still matter for migrations or compatibility.
How large is the GPT-6 context window?
GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna each document a 1,050,000-token context window and up to 128,000 output tokens.
How much does the GPT-6 API cost?
For Standard short-context text requests, Astra costs $10 per million input tokens and $50 per million output tokens; GPT-6.1 Sol costs $2 and $10; Luna costs $0.10 and $0.50. Cached input, cache writes and long-context rates differ.
When does long-context GPT-6 pricing apply?
When prompt input exceeds 272,000 tokens, OpenAI's published long-context schedule applies to the entire request: twice the input and cache rates and 1.5 times the output rate.
Which GPT-6 model should I use for coding?
Use GPT-6.1 Sol as a baseline for complex coding and agents, compare Astra on difficult or high-risk tasks, and test Luna on focused work. Verify cost per successful task using your own evaluation data.
Does GPT-6 support image input and audio output?
The family models compared here accept text and image input and produce text output. They are not dedicated speech, realtime audio or video-generation models.
Which GPT-6 models support none reasoning effort?
GPT-6 Luna and the original GPT-6 Sol support none. GPT-6 Astra and GPT-6.1 Sol support low, medium, high, xhigh and max but not none.
Can GPT-6 use Chat Completions?
Chat Completions supports model-specific requests, but Responses is the recommended API for reasoning with tools. Astra and GPT-6.1 Sol require Responses for tool calling; the original Sol and Luna have more limited Chat Completions tool support.
What changed between GPT-6 Sol and GPT-6.1 Sol?
GPT-6.1 Sol is a newer, more capable Sol model at the same Standard short-context $2 input and $10 output rates. Its cached input is $0.10 per million versus $0.20 for original GPT-6 Sol.
Are token prices the total cost of a GPT-6 agent?
Not necessarily. Total costs can include uncached input, cached reads, cache writes, output, long-context tier, tool calls, selected processing mode, regional charges, retries and operations.
Do all GPT-6 models have the same performance because they share 1.05M context?
No. The context and output numbers are serving limits, not quality scores. Compare factual accuracy, agent reliability, latency and downstream task outcomes across models.
PRIMARY DOCUMENTATION
Verify GPT-6 claims with OpenAI sources.
MODEL SELECTIONUsing GPT-6 and model migration ↗
API PRICINGOpenAI Standard and serving tiers ↗
MODEL CARDGPT-6 Astra specifications ↗
MODEL CARDGPT-6.1 Sol specifications ↗
MODEL CARDGPT-6 Luna specifications ↗
SXF DEEP DIVEOriginal GPT-6 family comparison ↗
SXF independently calculates the illustrative token-cost examples from provider-published rates. Editorial model-selection advice is not a universal accuracy ranking. The related-news stream and verified change ledger below remain visible and automatically updated.
LATEST DEVELOPMENT
GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents
OpenAI introduced GPT-6 Cyber on November 5, 2026 as a specialized security research model for authorized vulnerability discovery and validation. In its tests, it reports 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, and stronger adherence to assigned scope. Initial access is limited to approved Daybreak Red customers, so this is not a general ChatGPT model release.
Open latest signal ↗WHAT CHANGED
GPT-6: what changed.
No post-baseline factual changes recorded for GPT-6 Astra, GPT-6 Sol, GPT-6 Luna since 2026-09-27.
VERIFIED BASELINE2026-09-27Current facts remain aligned with the SXF ledger.
VERIFIED CHANGE HISTORY
GPT-6 verified history.
GPT-6 Sol
Baseline verified: 1,050,000 context · 128,000 max output · $2.00 input / $0.20 cached / $10.00 output per 1 million tokens
Official evidence ↗GPT-6 Luna
Baseline verified: 1,050,000 context · 128,000 max output · $0.10 input / $0.01 cached / $0.50 output per 1 million tokens
Official evidence ↗GPT-6 Astra
Baseline verified: 1,050,000 context · 128,000 max output · $10.00 input / $1.00 cached / $50.00 output per 1 million tokens
Official evidence ↗Append-only SXF ledger · 3 verified events · chain head 5cb61a0bb057…