OpenAI Just Slashed GPT-5.6 Prices 80% — Coding Agents Got Way Cheaper

OpenAI didn’t announce it with a keynote. They just updated the pricing page.

GPT-5.6 Luna: 80% cheaper.
GPT-5.6 Terra: 20% cheaper.
GPT-5.6 Sol: 2.5x faster with new Fast mode (at 2x price).

Effective July 30 for Luna and Terra, August 21 for Sol’s Fast mode. If you run coding agents — Cursor, Codex, Claude Code, OpenCode, Zed, any of them — your per-agent cost structure just shifted overnight.

The Numbers That Matter

Cached reads stay at 90% discount. Cache writes cost 1.25× base input (new for GPT-5.6+). Minimum cache life: 30 minutes.

Why This Breaks the Old Cost Models

Most teams budget for agents like this: model cost × tokens per call × calls per day.

That math just broke.

Luna at $0.20 / $1.20 is effectively “too cheap to meter” for agent workflows

At the old $1.00 / $6.00, a 500-call daily workflow at 3K tokens/call cost ~$9/day. At new pricing: $1.80/day. That’s $54/month → $2,600/year per developer saved.

Luna now undercuts GPT-5.4 nano ($0.20 / $1.25) while beating it on every agent benchmark. OpenAI’s own data shows Luna at 84.7% on Terminal-Bench 2.1 — ahead of GPT-5.5 (85.6%) at 1/25th the cost.

Terra at $2.00 / $12.00 is the new “default production model” for coding agents

Terra scores 87.4% on Terminal-Bench 2.1 vs GPT-5.5’s 85.6% — at 40% of GPT-5.5’s price. For teams that need reliability over raw speed, Terra is now the obvious default. You’d need a specific failure mode to justify Sol.

Sol Fast mode (2.5× speed, 2× price) changes the latency/cost tradeoff

When you need answers now — debugging a production outage, unblocking a deploy — Fast mode delivers Sol quality at 2.5× throughput. The per-task cost is similar (2× price / 2.5× speed = 0.8× cost per unit time), but the wall-clock difference matters when humans are waiting.

What This Means for Your Agent Stack

If you use Cursor (ChatGPT Work / Codex)

• Free/Go users: Terra is now your default — cheaper credits, same quota
• Plus/Pro/Enterprise: You can route routine work to Luna, complex work to Terra, fire-drills to Sol Fast
• Action: Check your Cursor model settings. Default may still be Sol.

If you use Codex CLI / ChatGPT Codex

• Same tier access as Cursor
• /fast flag in Codex now maps to Sol Fast mode
• Action: Update your AGENTS.md or team docs with the new routing logic

If you use Claude Code / Anthropic API

• Fable 5 still leads on long-context and instruction nuance (see our benchmark deep-dive)
• But the price gap widened: Fable 5 at ~$15/$75 vs Terra at $2/$12 is 7.5× more expensive
• Action: Keep Fable 5 for security-sensitive and long-context tasks. Move volume to Terra/Luna via OpenRouter or multi-provider setup.

If you use OpenCode / OpenRouter / BYOK setups

• OpenRouter already reflects new pricing
• Action: Re-run your cost projections. The “use the best model” advice is now “use the cheapest model that clears your quality bar”

If you self-host or use local models

• The bar for “worth paying for API” just moved significantly higher
• Luna at $0.20/$1.20 competes with GPU hour costs for small workloads
• Action: Recalculate your break-even point

The New Agent Routing Logic (Steal This)

# Recommended model routing for coding agent teams (August 2026)

routine_refactoring: gpt-5.6-luna      # $0.20/$1.20 — 84.7% Terminal-Bench
feature_development: gpt-5.6-terra     # $2.00/$12.00 — 87.4% Terminal-Bench
security_review:     claude-fable-5    # Premium — long-context, safety
production_debug:    gpt-5.6-sol-fast  # $10/$60 — 2.5× speed when blocked
architecture_design: gpt-5.6-sol       # $5/$30 — max reasoning, no rush

The rule: Default to Luna. Escalate to Terra when quality dips. Escalate to Sol/Sol Fast only when you hit a documented failure mode. Reserve Claude for the 5% of tasks where its specific strengths pay off.

Hidden Gotchas

1. Cache writes now cost 1.25× base input

For GPT-5.6+, writing a cacheable prefix costs 25% more. But a single cache read (90% discount) pays it back. Design your prompts with explicit cache breakpoints — OpenAI added cache_breakpoint support in the Responses API for this exact reason.

2. Fast mode is not “Sol but faster” — it’s a separate SKU

You opt in via processing_mode: "fast" in the API or /fast in Codex. It’s not automatic. If your agent framework doesn’t expose this, you’re leaving speed on the table.

3. Subscription credit burn changed

ChatGPT Work and Codex subscribers: Terra and Luna now consume fewer credits per call. Your quota lasts longer. But Sol Fast burns credits 2× faster. Monitor your usage dashboard.

4. The “Sol = GPT-5.5 price” trap

GPT-5.5 is still $5/$30. GPT-5.6 Sol is $5/$30. But GPT-5.6 Sol beats GPT-5.5 on every coding benchmark. There is zero reason to use GPT-5.5 for new deployments.

The Bottom Line

The era of “agent costs are negligible” just became real.

At Luna pricing, a 10-person team running 5,000 agent calls/day spends ~$18/day on model costs. That’s $540/month total — less than one senior engineer’s daily rate.

The bottleneck is no longer model pricing. It’s:

• Context management efficiency
• Prompt engineering quality
• Agent loop design (how many calls per task)
• Tool calling overhead

OpenAI just removed the last excuse for not running agents at scale.

What to Do Today

1. Audit your current model assignments — are you defaulting to Sol/Terra when Luna would work?
2. Update your team’s AGENTS.md / model routing config — bake in the new tier logic
3. Enable cache breakpoints in your agent prompts if you use Responses API
4. Test Sol Fast mode on your slowest agent workflows — measure wall-clock improvement
5. Re-run your monthly cost projection — you’ll likely cut 40-80% off the model line item

The pricing data in this article reflects OpenAI’s official API pricing page as of August 21, 2026. Rates can change without notice. Always verify at developers.openai.com/api/docs/pricing before committing spend.