One signal

Update What changed

GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock, giving developers an enterprise route to use OpenAI’s cheaper models through AWS. That materially extends the launch beyond the original announcement, which did not include access details.

AWS says Sol is aimed at recurring complex work such as coding, debugging, data analysis and multi-step tool use, while Luna is designed for high-volume classification, extraction, summaries and focused questions. Both support explicit prompt caching, allowing repeated instructions, tool definitions and reference material to be reused rather than processed from scratch on every request.

The Bedrock deployment also adds familiar enterprise plumbing: IAM access controls, CloudTrail audit logs, VPC endpoints through AWS PrivateLink and hardware-isolated inference. AWS says prompts and completions are not used to train models and that zero data retention can be requested through an account team. Its announcement also repeats OpenAI’s claim that Sol made roughly half as many factual mistakes as GPT-5.6 Sol on an internal evaluation. That result is company-reported, not an independent benchmark.

For buyers, the useful change is not simply another model badge.

Sources and evidence

• OpenAIDevs on X: GPT-6 Sol and Luna just landed in Astra’s orbit. Both launch today with API prices 50% lower than GPT-5.6. Build with Sol. Scale with Luna. To production and beyon: AWS has announced general availability of OpenAI’s GPT-6 Sol and GPT-6 Luna on Amazon Bedrock, with production controls including IAM, CloudTrail, PrivateLink, hardware-isolated inference and prompt caching; AWS also reports OpenAI’s internal factuality result for Sol.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s official Node SDK has added model identifiers for GPT-6 Sol and GPT-6 Luna in version 7.22.0, giving JavaScript developers a concrete route to target the two newly announced models from code.

The release note records the addition but does not establish access terms, pricing, performance or whether every API account can use the models. Those details remain separate from the SDK change. The practical point is narrower and useful: applications using openai-node can now refer to Sol and Luna through the library’s model catalogue rather than waiting for a later SDK release.

OpenAI’s developer account says the models launched at prices 50% lower than GPT-5.6, while the supplied release note itself does not state those prices. Treat that figure as OpenAI’s claim, not as an independently tested cost comparison. For developers, the next sensible check is whether their account, chosen endpoint and existing tooling support the new identifiers without further migration work.

Sources and evidence

• v7.22.0: OpenAI’s openai-node v7.22.0 adds GPT-6 Sol and GPT-6 Luna model identifiers; OpenAIDevs claims both launched at prices 50% lower than GPT-5.6.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s GPT-6 Sol and GPT-6 Luna appear to cost substantially less than their GPT-5.6 predecessors, giving developers a clearer sense of what the company’s cheaper model strategy means in practice. The reported prices put Sol at $2 per million input tokens and $10 per million output tokens, while Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens.

The Deep View reported those figures on 22 September, while CNBC separately reported that OpenAI had cut API prices by 50% against GPT-5.6 promotional pricing. Jordan Ligren also posted the same headline figures on X, although the supplied evidence does not establish an independent price comparison beyond those attributed accounts.

That makes the launch more useful to buyers than the original announcement alone, which described Sol and Luna as cheaper without giving prices. Sol is positioned for more complex work such as coding, while Luna is aimed at high-volume tasks including extraction and summarisation. The sensible next calculation is cost per completed task, not cost per token, because retries, longer outputs and review time have a habit of joining the invoice uninvited.

Sources and evidence

• jordanligren on X: GPT-6 Sol & Luna Just Made GPT-5.6 Hard to Justify Using GPT-6 Sol/Luna effectively gives you around 2x the subscription usage value. At this point, I don’t real: The Deep View and Jordan Ligren reported specific API prices for GPT-6 Sol and GPT-6 Luna, while CNBC reported 50% cuts against GPT-5.6 promotional pricing. The update adds concrete pricing to the existing OpenAI model announcement.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s GPT-6 Sol and GPT-6 Luna now have more concrete API details than the launch headline initially offered. Sol is positioned for complex coding and agentic workflows, while Luna is aimed at high-volume routine work such as extraction, classification and summarisation.

OpenAI’s model documentation says cached input tokens for the models cost 10% of the uncached rate. That could matter for applications repeatedly sending the same instructions, tool definitions or reference material, although the practical saving will depend on how much of each request can actually be cached.

The documentation also says the models are being integrated into ChatGPT, Codex, the desktop app and the ChatGPT API, with different rate limits across usage tiers. OpenAI says prices are 50% lower than the previous model versions. Those are the company’s figures, but they give developers a more useful starting point than the original promise of “cheaper” models. The sensible next calculation remains cost per completed task, not merely cost per token.

Sources and evidence

• GPT-6 Sol Model API Documentation: OpenAI’s GPT-6 Sol and GPT-6 Luna have documented roles, platform integrations, tiered rate limits and cached-input pricing, with the company saying token prices are 50% lower than previous versions.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s GPT-6 Sol and GPT-6 Luna have now acquired the detail that turns a cheaper-model announcement into something developers can actually price. VentureBeat reports that Luna costs $0.10 per million input tokens and $0.50 per million output tokens, while Sol costs $2 and $10 respectively. OpenAI says those are permanent prices, not introductory rates.

Sol is aimed at repeated complex work such as building features, reviewing code, debugging and data analysis. Luna is positioned for high-volume, more tightly defined jobs including summarisation, extraction and straightforward questions. Both are less capable than the flagship GPT-6 Astra, so the practical pitch is model routing: reserve the expensive brain for difficult jobs and send routine traffic to the cheaper options.

Against GPT-5.6, VentureBeat calculates that Sol is 50% cheaper for both input and output. Luna is 50% cheaper on input and 58.3% cheaper on output. Those figures make the launch meaningful for agent builders, but token prices are not the same thing as task costs. The useful test will be how often each model completes work successfully, how many retries it needs and whether the saving survives real workloads. The spreadsheet, having been invited to the party, is now in charge of the music.

Sources and evidence

• OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more - VentureBeat: VentureBeat reports that OpenAI launched GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 and $0.50 respectively, with Sol aimed at complex recurring work and Luna at high-volume routine tasks. The report says OpenAI described the prices as permanent and that they are materially lower than the GPT-5.6 predecessors.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

GitHub Copilot has added OpenAI’s GPT-6 Sol and GPT-6 Luna, extending the models beyond OpenAI’s own services and Amazon Bedrock into coding tools and integrated development environments. GitHub says both are available through usage-based billing across various Copilot plans.

Sol is positioned for interactive and agentic coding, while Luna is the lighter option for smaller, faster tasks. GitHub’s announcement says the models are available across the OpenAI API, ChatGPT, Codex and GitHub Copilot, giving developers another route to test them inside existing coding workflows rather than treating the launch as a distant model-card event.

The more useful number is Luna’s listed price: $0.10 per million input tokens and $0.50 per million output tokens. Sol is listed at $2 per million input tokens and $10 per million output tokens, with cached Sol input priced at 10% of the standard uncached rate. Those figures come from GitHub’s announcement, so teams should still calculate cost per completed task rather than admire the token tariff like a particularly nerdy menu.

Sources and evidence

• OpenAI's GPT-6 Sol and GPT-6 Luna now available - GitHub Changelog: GitHub says GPT-6 Sol and GPT-6 Luna are available in GitHub Copilot and related development environments, with Luna priced at $0.10 per million input tokens and $0.50 per million output tokens, and Sol at $2 per million input tokens and $10 per million output tokens.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

GPT-6 Sol and GPT-6 Luna are now generally available in Microsoft Foundry, giving Azure customers another route to OpenAI’s new models and a clearer split between demanding work and high-volume tasks.

Microsoft positions Sol as the general-purpose option for coding, complex knowledge work and multi-step agent workflows. Luna is the smaller, faster model for extraction, summarisation, request routing and routine customer interactions. Microsoft says both can be deployed through Standard plans across all 28 global regions, as well as US and EU data zones.

The Azure announcement also lists Provisioned Throughput for Astra and Sol, and Priority Processing for Sol in global regions and the US. It recommends choosing by cost per completed task rather than token price alone. That is sound advice, and pleasantly less glamorous than admiring a new model name like it is a limited-edition trainers drop.

Sources and evidence

• GPT-6 Astra, Sol, and Luna: For production agents in Microsoft Foundry - Microsoft Azure: Microsoft says GPT-6 Sol and GPT-6 Luna are generally available in Microsoft Foundry, with Sol aimed at general-purpose and complex agent work and Luna aimed at high-volume routine workloads.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

New comparison data gives OpenAI’s GPT-6 Sol and GPT-6 Luna launch a sharper edge: both models are reported to cost 50% less than their GPT-5.6 predecessors. Wccftech says Sol is also priced 50% below Anthropic’s Claude Opus 5.5 for input and output, while Luna undercuts the cited off-peak pricing for DeepSeek V4.1 Flash.

The performance picture is less tidy than the price tag. Artificial Analysis is cited as scoring Claude Opus 5.5 at 58 on its Intelligence Index, compared with 48 for GPT-6 Sol. The same account says Sol’s lower hallucination rate partly comes from refusing more often, reducing wrong answers by about a quarter while lowering accuracy to 54%. Luna’s accuracy is reported as broadly unchanged at 44%, despite also answering fewer questions.

The useful conclusion for developers is not simply “buy the cheaper model”. The new figures suggest a routing decision: Sol and Luna may reduce token and task costs, but refusal rates, shorter deliverables and regressions on some coding, computer-use and research-debugging tests could create extra retries or supervision. The bargain bin, as ever, has a returns policy written in benchmarks.

Sources and evidence

• OpenAI Unleashes A New Price War, With GPT-6 Sol And GPT-6 Luna Now Priced Below Claude Opus 5.5 And DeepSeek’s V4.1 Flash, Respectively, Negating The Rationale For Open-Weight Mod: Wccftech reports that GPT-6 Sol and GPT-6 Luna are 50% cheaper than their GPT-5.6 predecessors, while cited Artificial Analysis results show Sol trading lower cost and lower reported hallucination against refusal and accuracy trade-offs.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s GPT-6 Sol and GPT-6 Luna are now available in ChatGPT Work, Codex and the API, with Luna also coming to the desktop app and Free and Go plans. Sol is aimed at complex work such as coding, while Luna targets high-volume tasks with clear outcomes, including summarisation and information extraction.

TechCrunch reports that the new models cost half as much as their GPT-5.6 predecessors for API use. OpenAI also says GPT-6 Sol makes about half as many mistakes as the earlier Sol on its internal factuality evaluation, while claiming lower coding error rates. Those performance claims come from OpenAI, not an independent benchmark, so the price change is firmer news than the alleged leap in reliability.

The practical choice for developers is now model routing: reserve Sol for harder jobs and send routine traffic to Luna, rather than paying flagship rates for every request. OpenAI says the models are already available to most paid ChatGPT accounts through Work and Codex, with a gradual rollout across the main ChatGPT app and website. The useful calculation is cost per successfully completed task, because a cheap model that needs several retries has merely found a more affordable way to waste your afternoon.

Sources and evidence

• OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes - TechCrunch: OpenAI launched GPT-6 Sol and GPT-6 Luna as cheaper models for complex and high-volume workloads, with broad availability and company-reported reliability improvements.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI’s GPT-6 Sol and Luna now have more precise API prices: $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 respectively for Luna, according to 36Kr. That puts Luna firmly in the cheap, high-volume corner of the market, while Sol is positioned for more demanding coding and agent work.

36Kr also reports benchmark results that favour Sol on several professional and software-engineering tasks, including a reported 68.8% score on DeepSWE v1.1, compared with 69.9% for Claude Fable 5 and 74.2% for DeepSeek V4.1 Flash. The article says Sol’s cost per task was about 80% lower than Claude Fable 5’s in that test. These figures are reported by 36Kr from OpenAI’s launch material, not independently verified results, so they are useful pointers rather than a settled leaderboard.

The practical change for developers is clearer model routing. Luna may suit extraction, summaries, code review and other repeatable jobs, while Sol is the more plausible choice for complex coding and multi-step work. The relevant calculation is still cost per successfully completed task, including retries and human checking. Token prices are finally doing something useful beyond decorating a launch graphic, but the bill only matters if the work comes out right.

Sources and evidence

• GPT-6 New Model Launch: Disrupts AI Market, Challenges DeepSeek’s Dominance with Unbeatable Low Pricing - 36Kr: 36Kr reports that GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 and $0.50 respectively, alongside OpenAI-reported benchmark comparisons and lower claimed task costs. The pricing is attributable to the supplied report; the benchmark results are not independently verified.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic has launched Claude Opus 5.5 within hours of OpenAI’s GPT-6 Sol and GPT-6 Luna announcement, adding a significant new fact to the cheaper-model contest already covered here. Anthropic says Opus 5.5 performs at the level of its flagship Fable 5.1 for most work while costing 40% less to run than Opus 5 on typical workloads.

Anthropic says the savings will reach customers through price cuts and higher rate limits. It also describes Opus 5.5 as its most efficient model, with the release marking the first in a new family. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, according to Fortune’s account of the announcement.

The significance is bigger than another model-number parade. Both labs are pursuing everyday business work with cheaper alternatives while claiming to pass efficiency gains on to customers. That may make AI more affordable for companies watching their budgets, but it also puts pressure on the labs’ margins. Fortune quotes Ramp economist Ara Kharazian describing the market as a price war that could make models cheaper while making profitability harder to defend. The spreadsheet, as ever, would like a word.

Sources and evidence

• What AI slowdown? OpenAI, Anthropic release dueling models as price wars heat up - Fortune: Anthropic launched Claude Opus 5.5 alongside OpenAI’s GPT-6 Sol and GPT-6 Luna, saying Opus 5.5 matches Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. The releases sharpen competition around cheaper AI models for business users.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI says GPT-6 Sol nearly matches Claude Fable 5 on DeepSWE v1.1, a benchmark for long-horizon coding agents, at roughly 80% lower cost per task. The company also says GPT-6 Luna is comparable to Fable 5 at medium reasoning while costing 96% less per task.

The figures add a more useful layer to the GPT-6 Sol and Luna launch than token pricing alone. They suggest Sol is aimed at complex coding and agent work, while Luna is the cheaper option for repeatable, high-volume tasks. In practice, developers will need to compare the cost of a successfully completed job, including retries, tool calls and human checking, rather than admiring the API price in isolation.

The claims come from OpenAI’s comparison of DeepSWE v1.1 results and are not an independent benchmark. They are nevertheless material for teams deciding how to route workloads between OpenAI and competing models. A cheaper model that needs twice as many attempts has not discovered economics, merely renamed the invoice.

Sources and evidence

• OpenAIDevs on X: On DeepSWE v1.1, which tests coding agents on long-horizon engineering tasks, GPT-6 Sol (max) nearly matches Claude Fable 5 (xhigh) at ~80% lower cost per task. GP: OpenAI claims GPT-6 Sol nearly matches Claude Fable 5 on DeepSWE v1.1 at approximately 80% lower cost per task, and that GPT-6 Luna is comparable to Fable 5 at 96% lower cost per task.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

OpenAI says it has improved prompt caching in its GPT-6 API, with higher cache-hit rates by default for repeated input. The company says cached input tokens can receive discounts of up to 90%, while the cached context should also help agents respond faster.

The practical target is the repetitive material that agents keep sending: instructions, tool definitions and reference documents that remain largely unchanged between requests. More of that input qualifying for caching could reduce API bills without developers having to redesign every prompt, although the size of the saving will depend on cache hits and the workload’s request pattern.

This adds a useful operational detail to the existing GPT-6 Sol and Luna pricing picture. Token prices may be lower, but the real calculation for developers is now cost per completed task after caching, retries and latency, rather than the headline rate alone. OpenAI’s claim comes from its developer account and is not an independent measurement.

Sources and evidence

• OpenAIDevs on X: We’ve improved prompt caching in the API for GPT-6, helping agents run faster and cost less. Higher cache-hit rates by default mean more input tokens benefit from: OpenAI says improved GPT-6 API prompt caching will increase default cache-hit rates, allowing more repeated input tokens to receive discounts of up to 90% and helping agents run faster.

Independent WittyWires Watcher; not an official account or feed.