OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026. If your production workflow currently uses GPT-5.6 Sol or Luna, start a comparison pilot now. Keep the existing model in production until the new candidates have passed 30–50 anonymised cases from your actual workload, including the local languages your system must handle.
Teams with stable, validated Claude or other provider workflows can wait. OpenAI's launch announcement publishes lower prices and strong benchmark results, but neither shows how the models perform on Hungarian or other Central European documents, customer requests or internal tools. On the same day, AWS made both models generally available on Amazon Bedrock. Its announcement directs buyers to the service documentation for supported regions and Bedrock pricing.
For a broader procurement comparison, see the guide to proprietary enterprise AI models in autumn 2026. If Claude is also on your shortlist, a separate guide considers whether to switch to Opus 5.5.
The price cut affects workloads differently
OpenAI's price table shows Sol input pricing falling from $4 to $2 per million tokens, while output falls from $20 to $10. Luna input moves from $0.20 to $0.10, and output from $1.20 to $0.50.
| Item | GPT-5.6 promotional price | GPT-6 list price | Change |
|---|---|---|---|
| Sol input | $4 | $2 | 50% lower |
| Sol output | $20 | $10 | 50% lower |
| Luna input | $0.20 | $0.10 | 50% lower |
| Luna output | $1.20 | $0.50 | 58.3% lower |
All prices are per million tokens. Luna's output reduction is larger than the headline 50%, which may matter for output-heavy summarisation workloads.
The comparison uses GPT-5.6 promotional pricing. Compare the new rates with the price your company actually paid. Token rates also do not reveal total economics: Luna output is one twentieth of Sol output pricing, but the models may use different token volumes and achieve different success rates on the same job.
OpenAI also publishes a 90% discount for cached input-token reads and says GPT-6 should reuse prior context more often. That may materially affect agents with long, repeated prompts. The actual saving depends on cache use in your own requests, so it belongs in the pilot rather than the business case as an assumed discount.
Sol and Luna target different categories of work
OpenAI and AWS position Sol for recurring complex tasks, software development and agentic workflows that use multiple tools. Luna targets high-volume summarisation, extraction, classification and routing.
For a Hungarian document-processing assistant, Luna is a reasonable first candidate when the job is extracting short fields, assigning document categories or routing cases. The same may apply to summarising conversations and classifying incoming requests in a customer-service system. Sol merits a pilot when the workflow operates across several applications, performs multiple decision steps, changes code or validates a longer sequence of work.
The product tier alone does not determine the cheaper result. Run the same task definition, tools and acceptance criteria. Luna can still cost more if its results need additional manual correction or repeated attempts. If it meets the required quality, its much lower token rates can produce a meaningful saving.
Vendor benchmarks do not settle a Hungarian evaluation
OpenAI reports that GPT-6 Sol at xhigh effort scored 33.2% on AutomationBench at $0.27 per task. In the same table, Claude Opus 5 at max effort scored 26.9% and cost 11.1 times as much per task. That ratio is consistent with the vendor's statement that Sol delivered the higher score at about 9% of Opus 5's cost.
Derived from the vendor table, the implied Opus 5 cost is approximately $3.00 per AutomationBench task, compared with $0.27 for Sol. This calculation applies only to the benchmark conditions and cannot estimate the cost of a Hungarian or Central European business workflow.
The limitations matter. Competitor evaluation scores came from public reports, so the models were not all run in the same harness. The stated Claude Fable 5.1 cost excludes Opus 5 fallbacks even though fallback occurred on roughly 40% of tasks. OpenAI publishes these results and no independent party has replicated them, so do not rely on them without your own test.
For factuality, OpenAI reports that Sol made roughly half as many mistakes as GPT-5.6 Sol. The evaluation used conversations selected because users had previously flagged errors, and the company says the set is not representative of typical use. OpenAI's published results include no Hungarian-language or CEE-specific evaluation.
A local pilot should therefore define a successful case before testing. Count every retry, tool call, fallback and manual correction, then calculate cost per successful task. Replicating a benchmark score is less useful than establishing whether the model completes the actual business process at an acceptable quality and cost.
For model comparison before migration and ongoing production monitoring, review the managed operations service. Managed AI Operations
Existing GPT-5.6 users have the clearest reason to test
| Situation | Sensible decision | Reason |
|---|---|---|
| GPT-5.6 Sol or Luna handles substantial volume | Start a comparison pilot | List prices are lower and API model IDs are available |
| A validated Claude or other provider workflow is stable | Keep it for now | OpenAI's published comparison does not test Hungarian-language cases |
| Bedrock use is tied to a particular EU region | Check the documentation first | The AWS announcement does not list regions or pricing |
| The workflow is high-risk or tightly regulated | Complete your controls and contract review first | Vendor benchmarks do not validate your use case |
Migration testing should be relatively direct for an existing OpenAI API workflow because the new models are available as gpt-6-sol and gpt-6-luna. Pin the selected model identifier in production so a later default change does not silently alter behaviour.
In ChatGPT Enterprise, an administrator must enable the models. OpenAI's release notes say Sol and Luna are available in Work and Codex, but not in Chat.
Confirm the region and terms before using Bedrock
Bedrock provides another procurement and operating option alongside OpenAI's direct API. AWS says both models support up to one million tokens of context and can use established Bedrock controls for access, governance and auditing model invocations.
The 22 September announcement does not list supported regions or Bedrock pricing. A Hungarian or CEE buyer with an EU processing requirement should confirm the region, inference profile, price and data-processing terms before starting a pilot through Bedrock.
A one-million-token context window does not automatically make loading an entire document set economical. Compare retrieval-selected context, cached prefixes and full-document prompts on cost per successful task. A larger window is useful only when the added context improves the result enough to justify its processing cost.
Build the model comparison in five steps
- Select 30–50 real, anonymised cases from the current workflow. Include Hungarian cases, easy tasks, known exceptions and cases in every other language the system must handle.
- Define success for each task type: correct fields, an acceptable summary, accurate routing or valid tool use.
- Run the incumbent model, GPT-6 Sol and, where appropriate, Luna with the same prompt, tools and reasoning-effort setting. When comparing another model family, use the closest available effort level because vendor scales are not identical.
- Calculate cost per successful task using all tokens, cached reads, retries, fallbacks and manual correction. The AI ROI calculator can then compare the resulting operating cost with the current manual process.
- Before production rollout, confirm region and data terms, pin the model identifier and preserve a tested rollback path.