OpenAI Coverage

GPT-5.6 Luna: Free ChatGPT Gets Unlimited Text Chats for the First Time

OpenAI has replaced GPT-5.5 Instant with GPT-5.6 Luna as the default model for Free and Go users, removing text chat rate limits entirely. It is the first time a frontier-adjacent model has been made truly free at scale — and the pricing math behind it changes what developers should expect from "free AI."

Published August 20, 2026  ·  Source OpenAI announcement (August 6, 2026)  ·  Status Rolling out globally

01The Announcement: What Changed and When

On August 6, 2026, OpenAI announced a batch of changes to ChatGPT that collectively mark the most significant shift in free-tier access since the product launched. The headline: GPT-5.6 Luna becomes the default model for Free and Go users, replacing GPT-5.5 Instant. Text-based chats become unlimited — no rate limits, no message quotas, no throttling to a dumber model after hitting a cap. And a new Think button lets free users activate deeper reasoning on harder questions.

The rollout happened in two phases. GPT-5.6 Luna became the default model within the announcement week. Unlimited text chats and the Think button followed the next week, subject to abuse-prevention safeguards. File uploads, image generation, voice, data analysis, and other resource-intensive features remain subject to separate limits.

OpenAI framed the move in explicit terms: "Access shapes opportunity, and this update gives more people the ability to keep asking, develop an idea, and get help when they need it." The update arrived as ChatGPT crossed 1 billion weekly active users — a scale that makes even marginal per-user cost matter enormously.

02What GPT-5.6 Luna Actually Is

Luna is the lowest-cost tier in OpenAI's three-model GPT-5.6 family, sitting below Terra (mid-range) and Sol (flagship). The naming convention is deliberate: the number (5.6) identifies the generation, while Sol, Terra, and Luna are what OpenAI calls "durable capability tiers that can advance on their own cadence." GPT-5.6 Luna is the same model that powers the unified AI API at the lowest price point — $0.20 input / $1.20 output per million tokens after the July 30 price cut.

The performance profile is strong where it needs to be for everyday chat and reasonable where it can afford to be weak:

• Agents' Last Exam: 50.3% — outperforming Claude Fable 5 (40.5%) and Claude Opus 4.8 (45.2%) on long-horizon professional workflows.
• Terminal-Bench 2.1: 84.7% — competitive with Terra (87.4%) and ahead of Fable 5 (83.1%).
• DeepSWE v1.1: 67.2% — trailing Fable 5 (69.7%) by 2.5 points but at a fraction of the cost.
• GPQA Diamond: 92.3% — within 2 points of Sol (94.6%) and Fable 5 (92.6%).
• AA Intelligence Index: 51.2 — below Sol (58.9) and Fable 5 (59.9), but ahead of Gemini 3.5 Flash (50.2).

Where Luna falls short is also instructive. On MRCR long-context recall, Luna scores just 41.3% versus Sol's 91.5% and Terra's 89.6% — a cliff, not a gradual decline. This means Luna struggles with tasks that require retrieving information from very long documents or large codebases. For free-tier chat, where most conversations are short and self-contained, this weakness rarely surfaces. For developers considering Luna for production workloads, it is a hard ceiling.

03The Price That Changed Everything

The August 6 announcement built on a price cut that landed on July 30, 2026 — three weeks after GPT-5.6's general availability on July 9. OpenAI reduced Luna's API price by 80%, from $1.00/$6.00 to $0.20/$1.20 per million tokens (input/output). Terra dropped 20%. Sol stayed flat.

The stated reason was not competitive pressure but self-optimization: GPT-5.6 Sol autonomously rewrote and optimized production inference kernels, reducing end-to-end serving costs by 20% and increasing token-generation efficiency by more than 15%. OpenAI passed part of those savings downstream.

The resulting price landscape across the GPT-5.6 family and competitors:

Table 1 — Published API list prices as of August 2026. "vs Luna" = how many times more expensive per 1M-in + 1M-out. Real bills depend on input/output mix, caching, and context length.

The economic reality for free-tier users is even simpler: OpenAI is absorbing the cost. Luna's per-token API rate is real, but Free and Go users do not pay it. OpenAI is funding the gap between serving cost and zero revenue on roughly a billion weekly sessions — a bet that scale, lock-in, and upsell to Plus/Pro/Business justify the subsidy.

04Benchmark Breakdown: Luna vs the Frontier

The question is not whether Luna is as good as Sol. It is not, by design. The question is how much capability free users get for zero dollars, and where the gaps matter in practice.

Table 2 — GPT-5.6 Luna vs Terra, Sol, and leading competitors. Data from OpenAI's GPT-5.6 launch post (July 9, 2026) and Artificial Analysis. Vendor-reported where noted.

The pattern is clear: Luna is within 2–3 points of Terra and Sol on most benchmarks, but the gap widens on the hardest tasks (FrontierMath, long-context recall) and on agent-shaped work where token count matters. On Agents' Last Exam — arguably the most practically meaningful benchmark for knowledge work — Luna at 50.3% actually outperforms Claude Fable 5 (40.5%) and Opus 4.8 (45.2%). At $0.21 per task on the AA Intelligence Index, versus $1.04 for Sol, Luna delivers roughly five times the tasks per dollar.

05The Think Button: Luna's Missing Piece

The most interesting addition is not the model change — it is the Think button. Free and Go users now get a toggle that activates higher reasoning effort on GPT-5.6 Luna. For questions that require deeper analysis, multi-step logic, or domain expertise, Think gives Luna more compute time to work through the answer.

This matters because Luna's default reasoning effort is optimized for speed and cost, not depth. The Think button is OpenAI's way of letting free users opt into a higher-quality experience on-demand without upgrading plans — and without OpenAI paying for Sol-tier compute on every request. It is a pragmatic middle ground: most questions do not need deep reasoning, but when they do, the button is there.

"For questions that require deeper reasoning, free users can tap the new Think button to give GPT-5.6 Luna more time to work through the answer."

— OpenAI, August 6, 2026

For developers, Think is a signal about how inference cost will be managed going forward: most traffic stays cheap, and extra compute is activated only when the user explicitly requests it. The same pattern applies to API usage through the AICC model library, where teams can route based on task difficulty.

06What "Unlimited" Actually Means

Unlimited text chats means exactly what it says for text: no rate limits on text-based messages. You can send as many text prompts as you want, as many times as you want, without hitting a cap or being downgraded to a weaker model. Previously, Free and Go users were limited to roughly 10–40 text messages per 3–5 hours on flagship models before being throttled.

But "unlimited" has boundaries:

• File uploads — separate limits still apply.
• Image generation — still rate-limited.
• Voice mode — still rate-limited.
• Data analysis — still rate-limited.
• Abuse safeguards — automated abuse detection can impose limits; this is subject to OpenAI's standard terms.

The practical impact: text-based conversations — Q&A, writing assistance, brainstorming, coding questions, research queries — are now genuinely unlimited. For the majority of free-tier usage, this is a meaningful change. The resource-heavy features remain gated, which is reasonable given their compute cost.

07How This Compares to Other Free AI Tiers

OpenAI is not the only company offering free AI access. But the scale and model quality are different:

• Google Gemini — Gemini 3.5 Flash is free in the Gemini app, but with usage limits. Gemini 3.7 Flash costs $0.75/$3.75 through the API, which is 3.2× Luna's price. Google's free tier is limited; its API tier is not free.
• xAI Grok — Grok 4.6 is available through X Premium+ at $22/month. No free tier with comparable model quality. API pricing at $2/$6 for the lower context tier is 5.7× Luna.
• DeepSeek — DeepSeek V4 Pro is available at competitive API prices, but has no comparable free consumer product at billion-user scale. V4-Pro peak pricing increased sharply in August 2026.
• Anthropic Claude — Claude free tier exists with daily message limits. API pricing for Fable 5 at $10/$50 is 42.9× Luna.

What OpenAI has done is combine three things that no competitor has matched simultaneously: (1) a frontier-adjacent model (Luna), (2) unlimited usage at zero cost to the end user, and (3) a billion-user distribution surface. The economic moat is not the model — it is the install base and the habit.

08What This Means for Developers

For developers, the free-tier move has three implications that extend beyond ChatGPT:

1. The cost floor keeps dropping. Luna at $0.20/$1.20 — and effectively $0 for end users — resets expectations for what "cheap" means in AI. Teams that were paying $5/$30 for Sol-tier work on everyday tasks should audit their routing. A unified API gateway that can split traffic between Luna and Sol by task difficulty captures most of the quality at a fraction of the cost.

2. The Think pattern will spread. The idea of "cheap default, opt-in reasoning" is not unique to OpenAI — DeepSeek's V4-Pro pricing does something similar with peak/off-peak tiers. Expect every major provider to adopt difficulty-based routing as a standard pattern.

3. Free is the new trial. When a billion people can use GPT-5.6 Luna for free, the bar for charging for AI products rises. Applications that charge for basic chat or Q&A will struggle. Applications that charge for specialized workflows, domain expertise, or integrated tooling will survive. The enterprise plans that matter are the ones that solve problems Luna cannot.

09Frequently Asked Questions

OpenAI Coverage · www.ai.cc · August 2026