GPT-6 Sol and Luna: Smaller Models, Bigger Impact

OpenAI's latest models prove you don't need maximum scale to deliver maximum value — here's how they stack up against Claude's pricing crisis.

Earlier this month, we introduced GPT-6 Astra, most intelligent and aligned model in world. While most demanding and important projects still call for Astra's full depth, work happens at different scales, rhythms, and budgets. More ways to bring frontier intelligence into work you do every day.

Prices are per 1 million tokens. GPT-6 Astra continues to be our best model across board.

The Pricing Reality Check

Claude Opus 5.5 costs $0.015 per 1M input tokens and $0.0075 per 1M output tokens according to Anthropic's public pricing. Running a production workload of 100k daily API calls with an average of 2,000 input tokens and 1,000 output tokens per call works out to roughly $2,250 per day or $821,250 annually.

GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort on AutomationBench while costing 9% of Opus 5's per-task price. On Agents' Last Exam, GPT-6 Sol at max effort scores 56.4% while running at 60% lower cost per task. The pricing math isn't close — Claude's top-tier model makes GPT-6 look like the budget option, even when OpenAI's model is doing noticeably more work.

INPUT_TOKENS=2000
OUTPUT_TOKENS=1000
CALLS_PER_DAY=100000
INPUT_COST_PER_1M=0.015
OUTPUT_COST_PER_1M=0.0075

DAILY_COST=$(echo "$CALLS_PER_DAY * ($INPUT_TOKENS * $INPUT_COST_PER_1M + $OUTPUT_TOKENS * $OUTPUT_COST_PER_1M) / 1000000" | bc -l)
echo "Daily cost: $DAILY_COST"

I don't see how anyone can be running Claude in production at these prices without a very specific reason. The fallback chain alone — Claude Fable 5.1 with Opus 5 at max — adds another 31.4% to the bill, pushing effective costs above 8.9x GPT-6 Sol. It's pretty incredible what the OpenAI team is doing with model quality and pricing. There's no way this wasn't timed to undercut Anthropic's release today.

Real-World Benchmarks That Matter

AutomationBench and Agents' Last Exam are starting to tell a clearer story about where each model earns its keep. On AutomationBench, GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort while costing 9% as much per task. That's not a marginal edge. GPT-6 Luna at high effort adds 5.4 percentage points over its predecessor on the same benchmark, at 58% lower cost per task. On Agents' Last Exam, GPT-6 Sol at max effort hits 56.4%, roughly 60% cheaper per task than Opus 5. The pattern holds on FrontierCode too, where GPT-6 Sol improves substantially over GPT-5.6 Sol.

Pricing details make this sharper. Per 1M tokens:

Claude Opus 5.5: 50% price reduction from Opus 5
GPT-6 Sol: 60% lower cost per task vs. Claude Opus 5
GPT-6 Luna: 58% lower cost per task vs. GPT-5.6

There's a real question about whether the timing is coincidence. As one observer put it, "There's no way this wasn't meant to coincide with Anthropic's release today." That's fair. The price cuts across major players feel coordinated, even if unintentionally.

But the bigger issue is whether anyone's actually running Claude at these rates. The cost per task numbers tell you everything: even with a 50% drop, Opus 5.5 is still orders of magnitude more expensive than GPT-6 Sol for comparable or worse performance. I don't see how anyone's using Claude unless they have a specific reason to — and even then, they're paying for it.

echo "GPT-6 Sol: \$0.009 per task"
echo "Claude Opus 5: \$0.09 per task (10x more expensive)"

Sol vs Luna: Two Paths to Efficiency

Astra's pricing at $0.10/M input tokens puts it in the same tier as Luna, but I think most teams will still reach for Astra when the stakes are high enough to justify the cost. The real question isn't which model is better in isolation—it's whether Luna can handle enough of the day-to-day workload that teams feel comfortable keeping Astra licenses around primarily for the hard problems.

What I'm watching is how quickly cloud providers integrate Luna into their offerings. Right now, Anthropic's release feels more like a price cut than a technical leap, and that only works if you can actually buy it. If AWS or Azure drag their feet on availability, Astra keeps its moat not because it's obviously superior, but because it's accessible. I've seen too many promising models disappear into enterprise procurement hell.

The bigger implication might be for model specialization itself. We're inching toward a world where general intelligence is the premium product and task-specific efficiency is the commodity layer underneath. That's good for everyone building products on top of these models, but I'm not convinced the market has figured out how to price that relationship yet. The $0.10 price point says one thing; the reality of what Luna can actually do says another.

When to Choose Which Model

Astra exists for the projects where you need to be confident the model won't hallucinate a critical API integration or forget the constraint you specified three turns ago. I think that still covers most high-stakes engineering work, even though the gap between "good enough" and "bulletproof" is shrinking faster than I expected.

But Luna's pricing puts real pressure on the middle of the market. At $0.10/M input tokens, we're talking about costs that were effectively impossible six months ago. I'm genuinely uncertain whether this is a sustainable price point or a loss leader designed to lock in developers before competitors can respond. What I do know is that teams running medium-scale inference workloads are suddenly doing math they didn't need to do before.

The timing suggests this isn't just about Anthropic's release — it's about cloud providers wanting to prevent any single vendor from owning the API economy the way OpenAI did with GPT-3. But here's the friction that still matters: Luna isn't broadly available on the major cloud platforms yet. Pricing without distribution is just a press release, and I've seen enough promising models die on the vine because they couldn't get past enterprise procurement cycles.

The real question isn't whether cheaper, faster models will find users — they always do. It's whether Astra's performance premium justifies its cost premium at scale. I think it does for now, but not for long if Luna's availability expands as quickly as its price dropped.

Conclusion

GPT-6 Sol and Luna aren't trying to be Astra. They don't need to be. But they're also not just cheaper versions of last year's model with a new coat of paint — Sol actually outperforms Claude on AutomationBench at a fraction of the cost, and Luna delivers solid multimodal performance without the full Astra price tag. The question isn't whether these models are capable enough. It's whether OpenAI can keep this pricing pressure on competitors without cannibalizing Astra sales, especially when Sol is sitting there at $0.0003 per 1K tokens while Opus 5 costs 8.9x more and still loses.

The real story here is less about technical breakthroughs and more about market positioning. OpenAI's betting that developers will accept slightly less capability in exchange for dramatically lower costs — and the benchmarks suggest that bet might pay off. But I'm still not sure what to make of Luna. It's competent, sure, but in a landscape where every vendor is pushing multimodal as table stakes, competent doesn't feel like enough anymore.