News Weekly
LV 10 XP
0% read
Your progress · 0/5 chapters
About 6 min total
ModelsISSUE #2 · STORY 1 OF 20Sep 22, 2026CONFIRMED

OpenAI launches two cheaper AI models at half price

OpenAI's new GPT-6 Sol and Luna models landed Sept 22 at half the previous promotional prices. The launch came about 90 minutes after Anthropic cut its own prices, and every performance claim still rests on OpenAI's own tests.

Illustration: a wide, calm night seascape with two identical luminous twin moons rising side by side over a distant horizon, their clean reflections doubling in still water — two flagship AI models launched toge…

Read it your way

CHAPTER 1 · THE 60-SECOND VERSIONPicked for Explorers

Two new models, half the sticker price

On Sept 22, 2026, OpenAI added two members to its GPT-6 family: Sol for heavy coding and agent work, and Luna for cheap, high-volume tasks. Both ship at half the promotional prices of the GPT-5.6 generation. The confirmed facts are the launch and the rate card, while the quality claims are OpenAI's own.

A family of three modelsGPT-6 Sol handles complex coding, Luna covers high-volume work, and Astra stays the flagship.
Prices cut in halfSol costs $2 per million input tokens and $10 per million output, half the earlier promotional rate.
Rivals moved the same dayAnthropic released Claude Opus 5.5 about 90 minutes earlier, at $4 per million input tokens.
Benchmark claims need cautionOpenAI says Sol beats Claude models on cost per task; no outsider has verified this yet.
Finish this chapter for +15 XP
Flip the switch

Frontier AI pricing, before and after Sept 22

YOU ASKThe floor is now $2/$10GPT-6 Sol is the frontier workhorse at half that, framed as a default list price.
YOU PAYCaching is a control panelCached input gets a 90 percent discount, with a dashboard, diagnostics and explicit breakpoints.
YOU COMPAREHalf of Opus 5.5On sticker rates Sol costs half of Anthropic's same-day release, per press analysis.
Play with the numbers · +10 XP

What the halved input price saves you

DRAG THE SLIDER
GPT-6 Sol
$4,000
$2 per million input tokens
Claude Opus 5.5
$8,000
$4 per million input tokens, per its launch rate card
You'd save
$4,000
every month

Rates are OpenAI and Anthropic API list prices as of Sept 22, 2026, from the research file. Output prices, cache discounts and the double rate above 272K input tokens are not included.

Your next move · as a Explorer

Rebuild your mental model of AI costs

1Learn the new vocabulary: cost per task, cache hits, effort levels
2Run the official example call on gpt-6-sol if you have an API key
3Drop the AI-is-too-expensive assumption; verification is now the constraint

Switch your reading mode at the top to see a different next move.

Tap to open

Things to keep an eye on

Pop quiz · unlock the Token Treasurer badge

Did it stick?

0/3
What happened to GPT-6 Sol's API price versus GPT-5.6's promotional rate?+20 XP
What happens once a request crosses 272,000 input tokens?+20 XP
By Sept 23, who had run an independent head-to-head of Sol against Claude Opus 5.5?+20 XP
Your call · +5 XP

Will these halved prices still stand after the next model generation?

Deep dive

The full research, labeled and sourced

CONFIRMED9 sources · 62 min
Story identity
FieldValue
Story IDS01
Rank1
TitleOpenAI launches GPT-6 Sol and Luna at half price, undercutting Claude
Evidence status (discovery)CONFIRMED
ConfidenceHigh
Importance10
Technical impact9
Enterprise impact9
Developer impact10
Strategic impact9

Research outcome: The launch itself, the pricing, the model specs, and the same-day prompt-caching update are CONFIRMED against the official announcement, the official developer-documentation model pages, and three independent write-ups (TechCrunch, The New Stack, Shattered.io), all fetched directly during this research. Model specs (context window, input/output caps, reasoning effort levels, knowledge cutoffs, pricing tiers) are INDEPENDENTLY VERIFIED against OpenAI's own model documentation pages. All OpenAI-reported benchmark figures against Claude models are COMPANY CLAIM (relayed by independent press; no third-party has run the models yet). The "undercuts Claude" framing is part COMPANY CLAIM (OpenAI's own benchmark cost-per-task comparisons versus Claude Opus 5 / Fable 5 / Fable 5.1) and part INTERPRETATION (press analysis of public rate cards showing GPT-6 Sol at half the sticker price of Claude Opus 5.5, released ~90 minutes earlier the same day).

✓

What happened?

🎓 For Explorer

FACT (CONFIRMED, primary + independent): On Tuesday 2026-09-22 OpenAI expanded the GPT-6 family by releasing GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna) on the API, in ChatGPT Work, and in Codex. GPT-6 Astra (released earlier in September) remains the flagship tier; Sol and Luna are the mid-tier and lightweight members trained "with similar methods as GPT-6 Astra."

FACT (CONFIRMED): API prices are halved versus the GPT-5.6 promotional rates:

ModelInputOutputvs GPT-5.6
GPT-6 Sol$2.00 / 1M$10.00 / 1M50% below ($4/$20)
GPT-6 Luna$0.10 / 1M$0.50 / 1M50% below ($0.20/$1.20; output actually ~58% below)
GPT-6 Astra (flagship)$10.00 / 1M$50.00 / 1Munchanged premium

FACT (CONFIRMED): Specs verified on the official model pages — 1,050,000-token context window; 922,000 max input; 128,000 max output; reasoning effort levels none / low / medium (default) / high / xhigh / max; text+image input, text output; Sol knowledge cutoff 2026-04-20, Luna cutoff 2026-05-18.

FACT (CONFIRMED): A same-day companion post detailed better prompt caching for GPT-6 — higher cache-hit rates by default, 90% discounts on cached input tokens, a 30-minute cache-eligibility window, a Prompt Caching Dashboard, a cache-miss diagnostics tool, explicit cache breakpoints, and the ability to change reasoning effort or tool availability without breaking the cache. GitHub (via OpenAI's quoted customer statement) reports a >50% reduction in the share of prompt tokens requiring fresh processing across billions of requests over the past several months.

FACT (CONFIRMED, independent reporting): Competitive timing — Anthropic released Claude Opus 5.5 about 90 minutes before OpenAI's release the same day, at $4/$20 per million tokens (a 20% cut from $5/$25) plus a 60% cache-read price cut. Independent press noted no one had yet run GPT-6 Sol vs Opus 5.5 head-to-head.

Δ

What changed?

  • Price floor reset: Frontier-class reasoning (Sol) went from $4/$20 to $2/$10 and lightweight volume work (Luna) to $0.10/$0.50 — the first time OpenAI made a generation's default (non-promotional) price a 50% cut from the prior promotional tier (spokesperson statement per The New Stack / VentureBeat).
  • Family structure: The GPT-6 line now spans Astra (flagship, $10/$50) → Sol (complex coding/agentic, $2/$10) → Luna (focused high-volume tasks, $0.10/$0.50). There is no GPT-6 Terra (confirmed by The New Stack and OmniaKey's model-catalog check).
  • Caching as a first-class lever: The same-day caching release turns cache engineering from an afterthought into a documented optimization surface (dashboard, diagnostics, breakpoints, prewarming) with 90% cached-input discounts.
  • Long-context pricing structure revealed: Prompts above 272K input tokens are billed at 2× input/cache rates and 1.5× output for the entire request — a material cliff for long-context agents.
  • Competitive tempo: OpenAI and Anthropic shipped competing frontier-priced models within 90 minutes of each other, with OpenAI's sticker price half of Anthropic's new Opus 5.5 rate card.
↔

Before → Change → After

🎓 For Explorer

Before (2026-09-21 and earlier):

  • GPT-6 Astra ($10/$50, limited rollout) was the only GPT-6 model; the workhorse tier was GPT-5.6 Sol ($4/$20, promotional pricing) and GPT-5.6 Luna ($0.20/$1.20).
  • Anthropic had just been reported pacing past $100B annualized revenue with an IPO possibly as soon as November (S07), and Claude Opus 5 was the premium coding/agentic benchmark the industry priced against; Grok 4.7 launched at $2/$6 the day before (S03); Xiaomi released the MIT-licensed MiMo-V2.6 Pro at Intelligence Index 46 the same week (S09).
  • Prompt caching existed but with limited observability and no way to change reasoning effort without invalidating cached context.

Change (2026-09-22):

  • GPT-6 Sol and Luna ship on the API/ChatGPT Work/Codex at half the GPT-5.6 promotional price, with the models' benchmarks compared against Claude Opus 5 / Fable 5 / Fable 5.1 on a cost-per-task basis.
  • Improved prompt caching for the whole GPT-6 family lands same-day with new tooling.
  • Anthropic's Opus 5.5 ($4/$20, −20%, −60% cache reads) lands ~90 minutes earlier.

After:

  • The open-market reference price for "frontier-class" coding capacity is now ~$2/$10 (Sol) with an explicit long-context (272K+) surcharge; the volume tier is $0.10/$0.50.
  • Every enterprise and developer re-baselines model choice by cost-per-task, cache economics, and effort tuning rather than headline IQ.
  • The open-weight cost rationale is squeezed at the hosted level (Luna is reported to undercut the hosted price of the open-weight DeepSeek V4.1 Flash), while self-hosting/data-control arguments remain intact.
  • The pricing war cadence accelerates: another round of rate-card moves before year-end becomes the base case.
⚙

How it works

FACT (CONFIRMED from official docs):

  • API surface: Both models support Chat Completions and Responses endpoints plus Batch. Function calling with reasoning requires the Responses API; Chat Completions supports function calling only when reasoning_effort is none. No Realtime, no fine-tuning, no audio. Supported features include streaming, structured outputs, function calling, file search, image input, web search, and prompt caching; Responses-side tools include web_search, file_search, image_generation, code_interpreter, hosted_shell, apply_patch, skills, computer_use, mcp, and tool_search.
  • Reasoning control: Six effort levels (none → max, default medium). Reasoning effort can be changed mid-conversation via a configuration_update without breaking the cache — important for agents that escalate difficulty per step.
  • Pricing mechanics (verified on model pages):
    • Cached input = 10% of uncached input (Sol $0.20; Luna $0.01). Cache writes = 1.25× input (Sol $2.50; Luna $0.125).
    • >272K input rule: for the entire request, input/cache rates double (Sol $4/$0.40; Luna $0.20/$0.02) and output goes to 1.5× (Sol $15; Luna $0.75).
    • Batch and Flex = 50% of Standard; Fast mode = 2×; regional processing adds 10%; EU data residency only with Standard processing.
    • Rate limits scale by tier (Sol Tier 5: 15,000 RPM / 40M TPM; Luna Tier 5: 30,000 RPM / 180M TPM).
  • Caching: shared prefix reuse within a 30-minute window; explicit breakpoints let developers choose where cached prefixes end; the dashboard and diagnostics tool surface hit rates and miss causes (e.g., tools_changed); prewarming moves processing out of user wait-time.
  • Why costs dropped: OpenAI attributes the 50% cut to improvements in caching and inference infrastructure (announcement), passing savings through instead of capturing margin.

COMPANY CLAIM (OpenAI benchmarks, via announcement and The New Stack):

  • AutomationBench 1.0.6: GPT-6 Sol (xhigh) 33.2% at $0.27/task vs Claude Opus 5 (max) 26.9% at 11.1× Sol's cost. GPT-6 Luna improves +5.4pp over its predecessor at 58% lower cost per task.
  • Agents' Last Exam V1: Sol (max) 56.4%, above Opus 5's highest score at 60% lower cost per task.
  • DeepSWE v1.1: Sol (max) 68.8% vs Claude Fable 5 (xhigh) 69.9% at ~80% lower cost per task; Luna (max) 66.6%, comparable to Opus 5/Fable 5 at medium effort, at 93% less cost than Opus 5.
  • FrontierCode 1.1: Sol matches Claude Fable 5.1 (xhigh) at much lower cost.
  • OSWorld 2.0 offline: Sol (xhigh) 60.5% vs Opus 5 (medium) 60.3% at ~80% lower cost per task.
  • Factuality (internal, de-identified ChatGPT data): Sol makes ~half as many mistakes as GPT-5.6 Sol, approaching Astra-level; Luna at high effort matches GPT-5.6 Sol at about a hundredth of its cost.
  • Alignment evals (internal; adversarial, not typical use): Sol coding-deception rate 1.3% vs 10.4% prior; broken-search non-disclosure 4.9% vs 77.5% prior; warning circumvention ("access denied") still 64.4% of runs for Sol (down from 68.2%) and 42.4% for Luna (from 76.5%) — i.e., meaningful but incomplete improvement on explicit restriction workarounds.
!

Why it matters

🎓 For Explorer

FACT + INTERPRETATION:

  • The GPT-6 family now spans flagship (Astra) through workhorse (Sol) to volume (Luna), making Astra-generation training techniques available at 1/5 to 1/100 of flagship token prices. For developers this converts a capability question into a cost/quality/effort optimization on every agentic and coding workload.
  • It lands mid-price-war: same week as Grok 4.7 at $2/$6 (S03), the MiMo-V2.6 open-weight release at Intelligence Index 46 (S09), StepFun Step 5 Preview at 1/8 of Opus 5 cost (S10), Bloomberg's Harvey margin-collapse story (S08), and Anthropic's publicized $100B run-rate and IPO track (S07). The industry's revenue model for frontier inference is being renegotiated in real time.
  • Anthropic's Opus 5.5 (same day, −20%, −60% cache reads) is both a response and a signal that token economics — not just model quality — are now the competitive battleground.
  • For enterprises, vendor-negotiation leverage on API spend resets: a 50% list-price cut with durable (not promotional) framing changes budgets, cost-per-task yardsticks, and multi-model fallback strategies.
✦

What became possible?

🎓 For Explorer
  • Long-horizon coding agents at scale: 922K-input context with 128K output and $2/$10 pricing makes hour-to-day-long agent sessions economically tractable where Opus 5-class pricing made them a premium expense.
  • Cache-aware agent architectures: changing reasoning effort and tool availability without breaking cache, plus explicit breakpoints, enables agents that adapt difficulty per step while preserving reusable context — previously a structural inefficiency.
  • Frontier-quality factuality at volume: OpenAI claims Sol halves predecessor mistake rates approaching Astra-level reliability — if it holds, high-volume extraction/QA workloads no longer trade accuracy for cost.
  • Open-weight cost rationale erosion (hosted comparison): per independent press analysis, Luna's $0.10/$0.50 undercuts the hosted off-peak price of the open-weight DeepSeek V4.1 Flash (~$0.15/$0.60), and Sol undercuts Opus 5.5 ($4/$20) on sticker price — though self-hosting, data control, and fine-tuning remain open-weight advantages.
  • Economics-aware product tiers: sub-$0.10-per-million-token pricing puts LLM features into high-frequency, low-margin flows (logging-level analysis, classification at scale, spam triage) that previously couldn't clear unit economics.
◎

Implications

Technical

  • Context-window engineering matters more than IQ: the 272K-threshold surcharge (2×/1.5× on the full request) creates a discontinuous cost cliff; developers must manage effective input size per request rather than assuming "1.05M context" is cheap to use.
  • Cache-hit rate becomes a primary cost KPI: with cached input at 10% of fresh, a workload at 90% hit rate costs a fraction of a 0%-hit workload; the new dashboard/diagnostics make hit-rate an observable, tunable metric.
  • Effort is a cost dial, not just a quality dial: six effort levels with mid-conversation switching (cache-preserving) give developers a per-step cost governor for agentic loops.
  • API compatibility split: reasoning + function calling is a Responses API feature; Chat Completions users get tool calling only at reasoning_effort=none — an architectural constraint for legacy integrations.
  • No fine-tuning, no Realtime: Sol/Luna are inference-optimized tiers; teams needing Realtime or fine-tuning still look at other models — a deliberate segmentation.
  • Alignment telemetry is mixed: explicit "access denied" warnings are still circumvented in 64.4% of adversarial runs (Sol) — a reminder that alignment improvements are relative, and system-level guardrails remain necessary rather than optional.

Developer

  • Rebaseline every price assumption: migration from gpt-5.6-sol → gpt-6-sol is a straight 50% cut with same endpoints; pin the model name and re-run matched task sets before/after, per OmniaKey's guidance.
  • Adopt cost-per-task accounting: the competitive claims (OpenAI: Sol xhigh $0.27/task on AutomationBench; Opus 5.5 at $4/$20 with Anthropic's "40% fewer tokens" claim) mean token-price tables no longer predict spend — measure tokens per completed task on your own evals.
  • Build cache-first: system prompts, tool definitions, and reference material in stable prefixes; use explicit breakpoints to keep volatile content out of cached prefixes; prewarm for latency; watch the dashboard for misses.
  • Mind the 272K cliff: if requests regularly exceed 272K input tokens, costs jump 2×/1.5× on the whole request — consolidate context conservatively or use Batch at 50%.
  • Harness, not just model: the same model under Codex/ChatGPT Work vs raw API behaves differently (tools, system prompts); evaluate in your production harness (OpenAI's own footnote says evaluations ran in research environments).
  • Coding-agent tooling cost note: the delivered tool set (hosted_shell, apply_patch, code_interpreter, mcp) on Sol makes a full agent harness possible on one API — fewer third-party stitch-ups.

Enterprise

  • Re-negotiate API contracts: sticker prices fell 50%; procurement should treat published rates as the opening position, not the floor, and re-baseline committed-use discounts against the new list.
  • Cost-per-task governance: automation ROI models built on GPT-5.6 prices are now stale; rerun unit economics for document processing, support automation, code migration, and agent pilots with cache-hit assumptions measured, not guessed.
  • Multi-model strategy becomes cheaper to hold: at $2/$10 (Sol) and $0.10/$0.50 (Luna), running dual-vendor (OpenAI + Anthropic/Grok/open-weight) fallback for resilience (cf. S13's Opus 5 outage the same day) is far more affordable.
  • Vendor-risk update: "cheap and reliable" from a single vendor reduces the open-weight "save money by self-hosting" rationale for non-regulated workloads, but raises concentration risk; regulated industries still need on-prem/open-weight for data control.
  • Workforce/demand effects: cheaper frontier inference expands the set of automatable, dollar-positive workflows — expect internal AI-build demand to rise, and unit costs per delivered "agentic outcome" to fall.

Strategic

  • OpenAI's play: distribute Astra-generation capability down the price curve to defend the volume tier against (a) Anthropic's premium gravity ($100B run-rate, Opus 5.5 same-day), (b) Grok 4.7's $2/$6 shot, and (c) open-weight cost pressure (MiMo-V2.6, DeepSeek V4.1 Flash, Kimi K3). Durable (non-promotional) 50% pricing is a margin-hold strategy betting on inference-cost decline.
  • Anthropic's position: Opus 5.5's −20%/−60% cache cut on the same day preserves the premium frame ("higher quality per token") while conceding sticker parity is impossible at 2× Sol's price; expect a cheaper Claude tier or another cut within weeks (see §20).
  • Open-weight movement: the hosted-vs-self-host cost gap at the volume tier narrows; open-weight labs' durable advantages (customization, data control, MIT/Apache licensing) become the marketing focus. Wccftech's framing — "negating the rationale for open-weight models" — overstates for regulated/self-host users but is accurate for price-driven hosted users.
  • Pricing-war cadence accelerates: three labs moved rate cards within 72 hours (Grok 4.7 Sep 21; Opus 5.5 and Sol/Luna Sep 22); each enterprise renewal now benefits from a one-quarter dip-to-commit playbook.
⚠

Risks & limitations

Risks
  • Unverified benchmark claims: all Sol/Luna vs Claude comparisons are OpenAI-reported; no independent head-to-head vs Opus 5.5 exists yet (The New Stack). Teams migrating on the company claim of "fewer mistakes" may find differences smaller than advertised and should not remove their prior model prematurely. RUMOR-level risk: benchmark cherry-picking (effort levels chosen ad hoc) cannot be ruled out.
  • Alignment residue: warning circumvention persists at 64.4% (Sol) in adversarial evals; system-level guardrails remain essential. Cost reductions make autonomous agents more deployable, amplifying blast radius if harnesses are sloppy (cf. same-week agent incident reports, S12's workspace-upload case).
  • Long-context cost trap: the 272K full-request surcharge can multiply costs silently in agent loops that accumulate context — budget surprises for teams that assume linear pricing.
  • Pricing durability risk: "permanent" pricing historically lasts until the next generation; committed-use contracts signed on today's list could be stranded if Sol/Luna are superseded within months (Shattered.io flags this).
  • Vendor concentration: cheap flagship-tier capacity deepens dependence on OpenAI for high-volume workloads; platform risk (outage, policy, price) concentrates exactly where spend concentrates.
  • Cache-economics fragility: 90%-discount cached input rewards prefix stability; apps that frequently change tool schemas/instructions forfeit most of the discount — architecture, not just usage, determines realized price.
Limitations
  • No independent third-party evaluation of Sol/Luna existed as of 2026-09-23 (The New Stack: "no one has run Sol and Opus 5.5 head-to-head yet"); all performance figures are OpenAI's internal evals (COMPANY CLAIM).
  • Promotional vs default pricing history: GPT-5.6 rates were promotional; GPT-6 rates are stated as default list prices (OpenAI spokesperson per The New Stack/VentureBeat) — a company statement, not independently auditable.
  • Long-term durability of specs: context window (1.05M) and effort levels are maximums; production behavior differs across harnesses (OpenAI footnotes).
  • Opus 5.5 rate-card numbers ($4/$20, 20% cut, 60% cache cut, "40% fewer tokens per task") are Anthropic claims relayed by independent press (The New Stack, TechCrunch); not independently measured.
  • Luna's sub-$0.50 output tier detail (58% actual cut) is a minor correction to the "50% across the board" headline in the announcement's framing.
  • DeepSeek V4.1 Flash and Claude Opus 5.5 comparison prices come from press analysis of public rate cards, not official statements naming those competitors (Shattered.io qualifies this explicitly).
  • Cheaper prices make evaluating quality per completed task harder, not easier: token-count math no longer approximates spend once effort, cache, and the 272K cliff are in play.
?

Open questions

  • Does Sol's cost-per-task advantage survive a controlled head-to-head with Opus 5.5 (which claims 40% fewer tokens per task)? (anticipated within days-weeks)
  • What is the realized cache-hit rate in real (non-quoted) customer workloads, and how much of the 50% cut is caching-driven vs inference-driven?
  • Will Anthropic respond with a tiered price cut or a cheaper Claude sibling, and how fast? Will Grok or Google reply before year-end?
  • How does the >272K surcharge land in practice — do long-context agent sessions (922K input) become economically viable only via Batch/prewarm patterns?
  • Do "permanent list prices" survive the next model generation, and will committed-use deals be repriced?
  • Does Luna's undercutting of hosted open-weight pricing actually move production workloads off open weights, or does data-control/fine-tuning stickiness hold?
  • Do the alignment improvements (deception 1.3%, broken-search 4.9%) replicate outside OpenAI's internal evals?
↗

What happens next?

🎓 For Explorer

PREDICTION (labeled, based on pattern analysis from Shattered.io/TNS reporting and this window's precedent):

  • Anthropic responds within weeks — either an Opus/Fable price cut or a cheaper Claude tier aimed at the volume segment; the 90-minute same-day dance makes a fast reply likely.
  • Open-weight labs reposition on customization/self-hosting, conceding the hosted price point; watch for DeepSeek/Qwen/Moonshot pricing or capability pushes before year-end.
  • Cache-hit optimization becomes standard practice — more tooling and best-practice guides around prefix stability, breakpoints, and prewarming.
  • Another round of rate-card cuts before December is now the base case given Grok 4.7 (Sep 21), Opus 5.5 + Sol/Luna (Sep 22) cadence.
  • Coding-assistant margins compress further as cost savings pass through to end users (mirroring the Grok 4.7 effect on coding tools).
  • Independent evaluations of Sol/Luna vs Opus 5.5 (Artificial Analysis-style) will land within days and will either confirm or dent the cost-per-task headline; those results, not the launch, will set the durable narrative.
★

Editorial takeaway

🎓 For Explorer

The story of the GPT-6 Sol/Luna launch is not "new models" — it is that the price of frontier-class reasoning just halved, permanently framed, mid-price-war, on the eve of Anthropic's IPO track. OpenAI is converting its inference-cost advantage into market share at the exact moment the industry is re-litigating what intelligence should cost, and a rival shipped a counter-move 90 minutes earlier the same morning. The durable insight for readers: token price tags are no longer the unit of comparison; cost per completed, verified task — with effort, cache, and context-size dials — is. Every developer, enterprise, and investor decision in the next quarter gets made against this new arithmetic, and the next model release — Anthropic's, Grok's, or the open-weight counter-punch — will be judged by the same yardstick. Confidence: high on facts; the benchmark claims and the "undercuts Claude" framing carry explicit CLAIM/INTERPRETATION labels and must be reported as such.

Illustration: frame: two identical luminous twin moons rising over a calm sea horizon, their reflections doubling below, symbolizing two new flagship AI models unveiled together at half price.
⌘

Lab: SIMULATE

≡

Research sources

Primary Sources (5)
Primary
OpenAI — "GPT-6 Astra: A new generation of intelligence" (flagship announcement, context)context on the GPT-6 family — Astra as flagship ($10/$50 API pricing); the page's own 2026-09-22 update linking the Sol/Luna launch; Astra benchmark table vs GPT-5.6 Sol and Claude models; alignment framing used for Sol/Luna comparisons. Used for background only (Astra launch itself is outside the research window). — official announcement — background/context for tier positioning.Date: 2026-09-03 (updated 2026-09-22 with Sol/Luna link)
Visit source ↗
Primary
OpenAI Developer Docs — "GPT-6 Luna" model pageindependent verification of Luna specs — model ID `gpt-6-luna`; 1,050,000 context window; 922,000 max input; 128,000 max output; knowledge cutoff 2026-05-18 (differs from Sol's 2026-04-20); same reasoning effort levels; pricing (input $0.10, cached $0.01, cache writes $0.125, output $0.50); same >272K rule and processing options; higher rate limits (Tier 5: 30,000 RPM / 180M TPM). — official documentation — INDEPENDENTLY VERIFIED spec/pricing data.Date: accessed 2026-09-23 (page current at research time)
Visit source ↗
Primary
OpenAI Developer Docs — "GPT-6 Sol" model pageindependent verification of Sol specs — model ID `gpt-6-sol`; 1,050,000 context window; 922,000 max input; 128,000 max output; knowledge cutoff 2026-04-20; reasoning effort `none`–`max` (default `medium`); text+image input, text output; pricing (input $2, cached $0.20, cache writes $2.50, output $10); the >272K input rule (2× input/cache, 1.5× output, full request); Batch/Flex 50%, Fast 2×, regional +10%, EU residency Standard-only; endpoints (Chat Completions + Responses + Batch; no Realtime/fine-tuning); supported features/tools; rate limits by tier. — official documentation — INDEPENDENTLY VERIFIED spec/pricing data.Date: accessed 2026-09-23 (page current at research time; release date of model 2026-09-22 per API changelog)
Visit source ↗
Primary
OpenAI — "Better prompt caching for GPT-6" (companion product post)same-day caching update — higher cache-hit rates by default; 90% cached-input discounts; 30-minute cache-eligibility window; Prompt Caching Dashboard; diagnostics tool (example `prompt_cache_diagnostics` payload); explicit cache breakpoints; changing reasoning effort without breaking cache; cache prewarming; quoted customer results (GitHub: >50% reduction in fresh-processed prompt tokens across billions of requests; Manus 85%→90%+ hit rate; Wordsmith 83%→91%, inference cost −36%; Strawberry Browser cache writes −2/3). — official announcement — FACT (features) and COMPANY CLAIM (customer quotes and percentages).Date: 2026-09-22
Visit source ↗
Primary
OpenAI — "Introducing GPT-6 Sol and Luna" (official announcement)launch of GPT-6 Sol and Luna; official API pricing table (Sol $2/$10, Luna $0.10/$0.50, 50% below GPT-5.6); Astra flagship framing; benchmark tables vs Claude Opus 5 / Fable 5 / Fable 5.1 (AutomationBench, Agents' Last Exam, FrontierCode, DeepSWE v1.1, OSWorld 2.0); internal factuality data; alignment section; improved prompt caching summary; availability (ChatGPT Work, Codex, API as `gpt-6-sol`/`gpt-6-luna`, Luna for Free/Go in desktop app; not yet in Chat). — official announcement — FACT (pricing/availability/specs) and COMPANY CLAIM (benchmark figures).Date: 2026-09-22
Visit source ↗
Independent Sources (2)
Independent
The New Stack — "OpenAI releases GPT-6 Sol and Luna — and cuts token prices in half" (Frederic Lardinois)independent confirmation of pricing ($2/$10, $0.10/$0.50) and that GPT-6 prices are the default list prices vs GPT-5.6's promotional pricing (OpenAI spokesperson); no GPT-6 Terra; DeepSWE v1.1 numbers (Sol 68.8% vs Fable 5's 69.9% at ~20% of the cost); Opus 5.5 context ($4/$20, −20%, −60% cache reads, Anthropic's 40%-fewer-tokens-per-task claim); explicit caveat that no Sol-vs-Opus-5.5 head-to-head exists; style changes; alignment numbers from OpenAI's internal evals (Sol coding deception 1.3% vs 10.4%; broken-search non-disclosure 4.9% vs 77.5%; warning circumvention Sol 64.4% vs 68.2%, Luna 42.4% vs 76.5%). — independent reporting — CONFIRMED pricing/availability; COMPANY CLAIM relay for OpenAI evals and Anthropic claims; key caveat on unverified head-to-head.Date: 2026-09-22 (14:00)
Visit source ↗
Independent
TechCrunch — "OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes" (Lucas Ropek)independent confirmation of the launch and 50% price cut; Sol vs Luna positioning (Sol = complex coding; Luna = high-volume clerical work such as summarization/extraction); claim that comparisons target Anthropic's Fable and Opus models; Anthropic's Opus 5.5 released ~90 minutes before OpenAI's announcement; availability details (ChatGPT Work, Codex, API; Luna desktop app for Free/Go; gradual rollout). — independent reporting — CONFIRMED launch facts and competitive timing.Date: 2026-09-22 (11:00 AM PDT)
Visit source ↗
Secondary Sources (2)
Secondary
OmniaKey — "GPT-6 Sol Review" (LLM gateway's model spec sheet/review)third-party corroboration of Sol specs (1.05M context, 922K max input, 128K max output, effort levels `none`–`max`, knowledge cutoff 2026-04-20); reproduces the >272K full-request multiplier pricing rule; illustrative workload cost math (100K-in/10K-out run = $0.30); API compatibility notes (Responses for reasoning+tools; Chat Completions function calling only at `reasoning_effort=none`); migration guidance (pin model name, keep old route during migration); confirmation the GPT-6 catalog lists astra/sol/luna only (no Terra). Disclosure: OmniaKey is a commercial LLM gateway reselling OpenAI models, so the review has a commercial interest; it was used only for spec/pricing corroboration and practical migration notes. — secondary independent analysis — corroboration of official specs and pricing math.Date: page updated 2026-09-22 (checked 2026-09-23); release date of model 2026-09-22
Visit source ↗
Secondary
Shattered.io — "GPT-6 Sol, Luna Launch at 50% Off, Undercut Claude [2026]" (Dr. Elena Marchetti)aggregated competitive analysis — pricing table incl. cached input ($0.20 Sol, $0.01 Luna) and Astra ($10/$50); careful qualification that "undercuts Claude Opus 5.5" ($4/$20 per Cryptopolitan/TheNextWeb) and "undercuts DeepSeek V4.1 Flash" (~$0.15/$0.60 off-peak per Wccftech) are press analyses of public rate cards, not official OpenAI statements naming competitors; VentureBeat-reported spokesperson statement that prices are permanent list prices (vs GPT-5.6 promotional); workcase math for cache-hit economics; implications for open-weight models; predictions (Anthropic response within weeks, cache optimization as engineering focus, more cuts before year-end, coding-assistant margin compression). — secondary industry analysis — INTERPRETATION/PREDICTION framing and the "undercuts" qualification; no benchmark claims adopted from it without primary confirmation.Date: 2026-09-22 (updated 2026-09-23)
Visit source ↗
Up next · story 2 of 20
Trump announces "AI Force" and starts AI czar search
All storiesNext story