OpenAI has expanded the GPT-6 model line with the release of GPT-6 Sol and GPT-6 Luna. Following the debut of GPT-6 Astra earlier this month, the two new models translate Astra’s architectural improvements into lower latencies and reduced operational expenses. According to the official announcement from OpenAI, both models target agentic coding loops, high-frequency developer tasks, and enterprise automation.
Astra remains OpenAI’s flagship model for deep mathematical reasoning and high-stakes synthesis. Sol and Luna balance performance with compute efficiency, providing higher usage ceilings for teams running autonomous pipelines. OpenAI reduced API token rates by 50% compared to previous promotional tiers for GPT-5.6 models.
OpenAI expands the GPT-6 series with Sol and Luna, focused on cost-efficient coding and enterprise agents.
Token pricing and prompt cache architecture
Inference optimizations and updated caching routines cut serving expenses across both tiers. Input and output rates fell by half relative to GPT-5.6 Sol and Luna. In addition, prompt cache reads receive a 90% discount on recurring context tokens.
Model Tier
Input (per 1M tokens)
Output (per 1M tokens)
Cached Read Discount
Price Shift
GPT-6 Sol
$2.00
$10.00
90% off input
-50% vs GPT-5.6 Sol
GPT-6 Luna
$0.10
$0.50
90% off input
-50% vs GPT-5.6 Luna
GPT-6 Astra
Flagship rate
Flagship rate
Supported
Top capability baseline
API token rates and price reductions for the GPT-6 model family.
OpenAI introduced explicit cache breakpoints, letting developers control the exact boundaries of reused prompt prefixes. The engine also preserves cached context when users adjust reasoning effort or toggle available tools between steps. Telemetry published by GitHub showed that these caching updates cut un-cached prompt tokens by more than 50% across Copilot agent traffic.
Enterprise automation and professional knowledge benchmarks
On AutomationBench 1.0.6, hosted in collaboration with Zapier to evaluate agents across 47 real business tools, GPT-6 Sol at xhigh effort scored 33.2% accuracy at $0.27 per task. This mark surpassed Claude Opus 5 at max effort (26.9% at $3.00, roughly 11 times Sol’s cost) and Claude Fable 5.1 (31.4% at over 8.9 times Sol’s cost). GPT-6 Luna improved on GPT-5.6 Luna by 5.4 percentage points while cutting task cost by 58%.
AutomationBench evaluation comparing task completion accuracy against cost per task across frontier models.
On Agents’ Last Exam V1, which assesses long-horizon assignments across 55 commercial sectors, GPT-6 Sol at max effort scored 56.4%. That score exceeded Claude Opus 5’s top mark on the evaluation at 60% lower cost per assignment.
Agents’ Last Exam evaluation measuring multi-step professional work across 55 industry sectors.
Factual reliability and reduced hallucination rates
In internal tests based on real user conversations that triggered factual errors in earlier models, GPT-6 Sol cut mistake rates by approximately 50%. The model approached Astra-level accuracy while operating on a standard budget. GPT-6 Luna matched the factual reliability of GPT-5.6 Sol at roughly one-hundredth of the compute cost.
Internal evaluation measuring error reductions in de-identified conversations with previous factual mistakes.
Software engineering, mergeability, and coding benchmarks
On FrontierCode 1.1 Main, which tests pull requests for both functional correctness and codebase mergeability (code style, scope restraint, test quality), GPT-6 Sol matched Claude Fable 5.1 at xhigh effort with lower token overhead. The result reflects tighter adherence to existing project conventions during multi-file edits.
FrontierCode 1.1 evaluation measuring pull request correctness and production mergeability.
On DeepSWE 1.1, an evaluation focused on multi-file issue resolution in real production repositories, GPT-6 Sol at max effort scored 68.8%. This placed the model within 1.1 percentage points of Claude Fable 5’s peak score (69.9%) at approximately 80% lower cost per solved issue. GPT-6 Luna at max effort reached 66.6%, matching Claude Opus 5 and Fable 5 at medium effort while cutting task costs by 93% and 96%.
DeepSWE 1.1 benchmark measuring complex bug resolution and feature additions in production codebases.
Benchmark
GPT-6 Sol
GPT-6 Luna
GPT-6 Astra
Claude Opus 5
Claude Fable 5.1
AutomationBench 1.0.6
33.2% ($0.27)
20.8% ($0.05)
41.4%
26.9% ($3.00)
31.4% (>$2.40)
Agents’ Last Exam V1
56.4% (max)
42.1%
63.2%
52.8%
(not reported)
DeepSWE 1.1
68.8% (max)
66.6% (max)
73.5%
65.4% (med)
69.9% (xhigh)
OSWorld 2.0 (Offline)
60.5% (xhigh)
45.2% (max)
71.0%
60.3% (med)
(not reported)
Comparative evaluation scores across agentic automation, coding, and computer use.
Computer use and autonomous GUI operations
On the OSWorld 2.0 offline benchmark, which tracks desktop operating system automation on complex workflows, GPT-6 Sol at xhigh effort scored 60.5% partial task completion. This matched Claude Opus 5 at medium effort (60.3%) at approximately 80% lower cost per run. GPT-6 Luna at max effort outscored GPT-5.6 Sol at medium effort at a tenth of its operating expense.
OSWorld 2.0 computer-use benchmark evaluating partial task completion across desktop applications.
Communication style and concise outputs
OpenAI applied Astra’s updated communication profile to Sol and Luna. In practical scenarios such as web UI component redesigns, GPT-6 Sol generates shorter answers, avoids restating obvious parameters, provides clear verification summaries, and avoids exposing unnecessary background prompts.
Side-by-side prompt output comparison showing direct, concise explanations in GPT-6 Sol.
Safety, alignment checks, and coding honesty
In alignment evaluations designed to test model honesty under pressure, Sol and Luna registered fewer misleading claims about completed code. In specialized tests measuring reviewer bypass, warning circumvention, and simulated deception, both models recorded lower infraction rates than previous releases.
Internal evaluation measuring coding honesty, warning circumvention, and reviewer bypass across model generations.
Availability and deployment timeline
GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Education accounts. Free and Go tier users have access to GPT-6 Luna through the official desktop application. In the OpenAI API, developers can access the models under the identifiers gpt-6-sol and gpt-6-luna with rolling rollout through the day.
Alibaba Releases Qwen3.8-Omni-Flash Alibaba has introduced Qwen3.8-Omni-Flash, an omnimodal AI model designed for low-latency visual, audio, and language…
