OpenAI Cuts GPT-5.6 Luna 80% and Terra 20%: The Cost-Per-Task Math for Agent Builders

Three weeks after shipping the GPT-5.6 family, OpenAI cut its prices. GPT-5.6 Luna is down 80%, GPT-5.6 Terra is down 20%, and GPT-5.6 Sol gains a faster paid option in the API. The lower Luna and Terra prices also flow through to how usage is counted in Codex and ChatGPT Work, so subscription quotas stretch further without any plan change.

The headline number is Luna. Per OpenAI's model documentation, Luna now costs $0.20 per million input tokens, $0.02 per million cached input tokens, and $1.20 per million output tokens. At launch on July 9 it was $1 input and $6 output. Terra moves from $2.50/$15 to $2/$12 per million tokens, per CNBC. Sol's standard pricing is unchanged at $5/$30.

The new price sheet#

Luna keeps its 1,050,000-token context window and 128K max output at the new price. That combination - million-token context at $0.20 input with $0.02 cache reads - is the part agent builders should stare at, because agents are input-heavy by construction: every tool call replays the conversation so far.

OpenAI's own chart makes the argument#

OpenAI's announcement leaned on Artificial Analysis data rather than adjectives:

Chart by OpenAI (via Artificial Analysis), from the announcement post on X. Intelligence Index v4.1 plotted against cost per task.

Read the chart the way OpenAI wants you to: Luna lands around a 51 on the Intelligence Index at roughly $0.05 per task, while GLM-5.2 Max, Claude Opus 5 (low effort), and Gemini 3.6 Flash sit at comparable intelligence but 5-10x the cost per task. DeepSeek V4 Pro is cheaper-adjacent but scores around 44. The claim is not "we are the smartest" - Sol still carries that flag at unchanged prices. The claim is "we now own the efficiency frontier," which is exactly the ground DeepSeek and GLM had been winning on. A CNBC report earlier in July found Chinese models had captured 46% of US enterprise token usage on OpenRouter; an 80% cut on the efficiency tier is a direct response to that graph, not a coincidence of timing. We walked through why cost per task beats cost per token as the deciding metric in our Fable 5 pricing analysis, and this chart is OpenAI making the same argument in the other direction.

Cost-per-task math for agent workloads#

Take a mid-sized agent task: 40 tool-calling turns, growing context, call it 2M cumulative input tokens (75% cache-hittable after the first few turns) and 60K output tokens.

At launch pricing: 0.5M fresh input at $1 ($0.50) + 1.5M cached at $0.10 ($0.15) + 60K output at $6 ($0.36) = about $1.01 per task.

At new pricing: 0.5M fresh input at $0.20 ($0.10) + 1.5M cached at $0.02 ($0.03) + 60K output at $1.20 ($0.072) = about $0.20 per task.

Same task, one-fifth the cost. Run that agent 10,000 times a month and the bill drops from roughly $10,100 to $2,000. That moves whole categories of workloads - triage bots, doc-wide refactors, evaluation harnesses, subagent fan-outs - from "meter it carefully" to "leave it running." It is the strongest counterexample yet to the AI affordability crisis worry that agent costs only ratchet upward.

The architecture implication is bigger than the bill. The standard cost-control pattern has been router-heavy: send easy calls to a cheap model, escalate hard ones. At $0.20/$1.20, the escalation threshold moves. A Luna-first fleet with Terra or Sol reserved for planning steps is now defensible where a month ago it looked stingy, and multi-agent designs like the ones in our parallel agent fan-out writeup get 5x cheaper on the worker tier where most tokens are burned.

Codex and ChatGPT Work: quotas quietly got bigger#

The second half of the announcement matters to anyone on a Codex or ChatGPT Work plan: the lower Luna and Terra prices are reflected in how usage is counted. Subscription prices and quota budgets stay the same, but Terra and Luna calls now consume proportionally less of them. In practice that is a silent capacity increase - the same $20 or $200 plan runs meaningfully more Luna-tier agent work per month. If you have been rationing Codex background tasks against rate limits, re-test your assumptions; our Codex developer guide covers where those limits bite.

Sol's change goes the other direction. Fast mode replaces Priority Processing in the API, delivering up to 2.5x faster responses at 2x the standard Sol price, per OpenAI's Fast mode docs; existing priority-tagged requests migrate automatically, and it aligns with /fast in Codex. That is a segmentation play: pay less for bulk intelligence, pay more for latency. For interactive agent UIs where the human is watching the stream, 2.5x faster at 2x price can be worth it; for background queues it obviously is not.

What it pressures competitors on#

The pricing spread this leaves across the market is stark. Luna at $0.20/$1.20 undercuts Gemini 3.6 Flash and Claude's Haiku-class models on list price while benchmarking above them on the Artificial Analysis index, and it takes away the primary reason enterprises were routing to DeepSeek V4 and GLM-5.2: price. Those labs still hold an open-weights card OpenAI cannot match, but "half the cost" was the easier procurement argument, and it just evaporated.

For Anthropic and Google, the pressure point is the efficiency tier, not the frontier. Sol staying at $5/$30 says OpenAI does not feel price pressure at the top - Opus 5 and Fable 5 compete there on capability. But every fleet design now benchmarks its cheap tier against a $0.05-per-task Luna, and our pricing landscape tracker has fresh numbers to absorb. Expect responses within weeks, not quarters; this market has never let an 80% cut sit unanswered.

One caution: list price is not cost per task. Luna's score on the index is a benchmark aggregate, and if Luna takes more retries or more turns than a stronger model on your workload, the 5x sticker advantage shrinks. The only numbers that settle it are your own evals with your own token traces. But the direction is unambiguous - the floor for capable agent intelligence dropped 80% overnight, and every budget spreadsheet built before July 30 is now wrong in your favor.

Sources#

• OpenAI announcement on X - the primary announcement, including the Artificial Analysis chart
• Advancing the price-performance frontier with GPT-5.6 - OpenAI
• GPT-5.6 Luna model docs - OpenAI - current per-token prices
• Fast mode - OpenAI API docs
• OpenAI API pricing
• CNBC: OpenAI cuts prices for two of its GPT-5.6 AI models
• VentureBeat: AI price wars

Continue Reading#

• Claude Fable 5 Pricing: Real Cost Per Task - the cost-per-outcome framework applied to Anthropic's frontier tier
• The AI Affordability Crisis - why agent costs were trending the other way before this cut
• AI Coding Tools Pricing 2026 - the full subscription and API pricing landscape
• Codex in 2026: The Developer Guide - where Codex quotas and rate limits actually bite
• Parallel Agent Fan-Out Day - the multi-agent patterns that get 5x cheaper under Luna's new price