GPT-5.6 Sol vs Claude Fable 5: OpenAI's Price Squeeze on Anthropic

OpenAI's new flagship matches Claude Fable 5 on intelligence at roughly a third of the cost per task. The real question is whether that price is a subsidy or a strategy.

OpenAI's new flagship, GPT-5.6 Sol, roughly matches Anthropic's Claude Fable 5 on raw intelligence while costing about a third as much to run the same task. That single fact is the whole story of this launch. It quietly moves the AI model race from "who is smartest" to "who is smart enough, for the least money," and that is the shift most worth understanding if you are trying to figure out where the profits in AI eventually land.

Near-parity on intelligence, then a big undercut

On the Artificial Analysis Intelligence Index, a composite score across many tests, Sol lands at 59 to Fable 5's 60. That is a rounding error, not a gap. The difference shows up on the bill. At maximum reasoning effort, Sol costs about $1.04 to complete one benchmark task versus roughly $2.75 for Fable 5, close to a third of the price. On raw API rates, Sol is $5 per million input tokens and $30 per million output tokens, exactly half of Fable 5's $10 and $50.

To put "a million tokens" in human terms: it is very roughly 700,000 words, about nine copies of a full-length novel. OpenAI is charging half as much to read and write that pile of text, at nearly the same quality.

A three-model line aimed at every price point

OpenAI did not ship one model, it shipped three: Sol, the flagship, Terra, a mid-tier built for high-volume work, and Luna, a cheap everyday model. Terra runs roughly 50% below Sol on cost per task and Luna about 80% below. OpenAI claims Terra and Luna still beat Fable 5 on its Agents' Last Exam test at around one-sixteenth of Fable's cost.

The strategic point is coverage. Instead of fighting Anthropic only at the premium top, OpenAI now has a product at nearly every rung of the price ladder. There is an efficiency layer underneath the sticker price too: OpenAI says Sol burns up to 54% fewer output tokens on agentic coding work than similarly capable models, and cache reads get a 90% discount. Because you pay per token, fewer tokens means the real-world gap can be wider than the headline rates suggest.

The real battle is agentic coding, and the scorecard is split

The money in enterprise AI increasingly sits in "agentic" work, where a model plans, calls tools, and grinds through multi-step jobs on its own. That is exactly where the two models trade blows.

Sol leads on the agentic tests. On Terminal-Bench 2.1, agentic terminal work, it scores 88.8, rising to 91.9 in a four-agent "Ultra" mode, versus Fable 5's 83.4. On Agents' Last Exam it posts 53.6 to Fable's 40.5. On the Artificial Analysis Coding Agent Index it edges ahead, 80.0 to 77.2. But on SWE-Bench Pro, which measures end-to-end code fixes on real repositories, Fable 5 wins clearly, 80.3 to 64.6. OpenAI disputes that particular benchmark, which itself tells you how much the framing matters.

The honest read: Sol wins the orchestration and long-horizon agent tests, Fable still wins the "actually resolve this bug" test. Which one matters depends on the workload.

The question hanging over the price: subsidy or strategy?

A cheap price only pressures a rival if it is sustainable. Nobody outside the two labs knows whether Sol's economics are real or subsidized. The AI Explained channel, reviewing the launch, called the pricing "a tell," arguing OpenAI is "gunning for anyone looking to save a buck versus Claude," and asked the question directly: is this "a last gambit to take market share from Anthropic, or a sustainable lower pricing?"

That is the crux. If the low price reflects genuinely cheaper unit costs, Anthropic has a structural problem. If it is a loss leader bought to grab share, Anthropic can afford to wait it out. The same review noted a sharpening detail: as bundled subscription access to Fable 5 gets pared back, more users feel the full API rate, which widens the very gap OpenAI is exploiting.

Anthropic's answer: quality and cadence, not price

Anthropic is not matching the price. It is defending on capability. The consensus among testers is that Fable 5 still has the better "judgment," a stronger feel on ambiguous, open-ended work, plus that clear SWE-Bench Pro lead on real code fixes. Early tester Ethan Mollick described them as similar in ability but different in temperament: Fable "wants to go off and do work on its own pace," while Sol "is faster but works with you in steps more."

On cadence, one closely followed model watcher, @zephyr_z9, argued Anthropic is "sitting on the throne and is clearly ahead, already post-training and getting ready for Fable 5.1." The bet there is that fast iteration and a quality edge keep the highest-value workloads even if Anthropic cedes the price-sensitive volume.

The number that actually matters: share, not crowns

Here is the uncomfortable data point for Anthropic. Benchmark crowns have not been converting into usage. SemiAnalysis, a respected chip-and-compute research shop, reported that OpenAI's usage share kept growing versus Anthropic even right after Anthropic's own frontier launch. In plain terms, you can hold the "smartest model" title and still lose the volume war to a rival with better distribution and a lower price.

That reframes the whole contest. The open question, raised on the Moonshots podcast, is whether a benchmark lead converts into enterprise contract wins before the competitor responds on price. With Sol, OpenAI just responded.

The bear case on Sol's own story

The counter-argument is worth holding onto. Model tester @synthwavedd cautioned that Sol tends to clear Fable only when maxed out in the multi-agent "Ultra" configuration, which he warned can "burn an ungodly amount of tokens," and that OpenAI was selective about which benchmarks it published. If real workloads need the expensive Ultra mode to match Fable, the clean "one-third the cost" story narrows on the tasks that actually pay. The sticker discount is real; the discount per unit of delivered quality is fuzzier.

What to watch

A few signals will settle which reading is right:

• Whether OpenAI's pricing holds through the next quarter or quietly rises, the clearest tell on subsidy versus sustainable cost.
• Usage-share data from independent shops like SemiAnalysis, which measures adoption rather than benchmarks.
• Anthropic's move: a Fable 5 price cut would concede the pressure, a fast Fable 5.1 would double down on the quality bet.
• Enterprise agentic-coding wins, the workload where the real budgets sit and where the split scorecard makes the outcome genuinely uncertain.

The tidy summary: OpenAI is betting that "smart enough, and much cheaper, at every tier" beats "smartest and priciest." Anthropic is betting that judgment and iteration speed hold the workloads that matter. The winner gets decided by whether Sol's price is a weapon OpenAI can afford to keep firing.

Sources

• @OpenAI — Sol vs Fable 5 on Agents' Last Exam and relative cost
• @ArtificialAnlys — Intelligence Index scores and cost per task
• The Decoder — cost per task, Terra/Luna tiers, token efficiency
• ZDNet — OpenAI's price/speed/productivity positioning
• @tengyanAI — agentic and coding benchmark breakdown
• @emollick — early-tester impressions of both models
• @zephyr_z9 — Anthropic positioning and Fable 5.1
• @synthwavedd — bear case on Sol's Ultra-mode economics
• AI Explained — pricing "tell" and subsidy question

More research at bargo.ai/research.