When Two AI Labs Handed Their Best Models a Dugout — and We Turned a Knob
Anthropic vs. OpenAI — and a Reasoning Dial We Could Turn
The 2026 World Series, five games, two AI managers from rival labs. Claude Fable 5 calling shots for the Dodgers. GPT-5.6 Sol running the Yankees. And on Sol’s side, a dial we could turn from “medium” to “high” to make it deliberate harder on every decision.
We ran it twice — identical rosters, same random seed, same everything — except Sol’s reasoning effort level changed between runs. The Dodgers won 4-1 in five games both times. What actually changed, and what that tells us about frontier reasoning models making real-time, messy, open-ended calls: that’s the story.
The Setup — and the Two Very Different Questions Inside It
Before anything else, a methodological split that matters.
There are two comparisons here, and they have completely different evidential weight.
The Fable 5 vs. Sol cross-model comparison is a case study, not a benchmark. The Dodgers have a different roster than the Yankees. Fable 5 is running “The Optimizer” personality (Dave Roberts — data-driven, platoon-heavy, aggressive bullpen management) while Sol runs “The Pressure Cooker” (Aaron Boone — win-now, veteran-trust, fiery). They’re from different companies at different price points. If the Dodgers win, you cannot conclude Fable 5 is the better model. The teams aren’t the same team. What you can do is examine the texture of each model’s reasoning — how it frames decisions, where its confidence sits, what logic it reaches for under pressure.
The Sol medium vs. Sol high comparison is clean. Same team, same roster, same personality prompt, same random seed, same opponent. The only changed variable is how hard Sol was instructed to think. Any difference here — in decisions made, confidence expressed, tokens spent, or outcomes achieved — is attributable to reasoning effort alone. This is the axis where real claims live.
How does “reasoning effort” actually work? Both models do hidden thinking before producing an answer — a scratchpad process invisible in the final output, but billed like any other tokens. Fable 5 reasons by default; you tune its depth via an effort setting, and here it ran at default. Sol reasons on demand, with an explicit low/medium/high dial. Turning that dial up tells Sol to spend more of its scratchpad working through the problem before committing to an answer. Critically, those hidden reasoning tokens are billed at the output rate — on Sol, that’s $30 per million tokens. More thinking isn’t free.
The Head-to-Head Texture: What Each Model’s Reasoning Actually Sounds Like
Because the cross-model axis is confounded, look past the scoreboard and into the decision logs.
Fable 5 reasons in frameworks. Its pull-pitcher calls almost always invoke the same three variables explicitly — batters faced, pitch count relative to a 30-pitch reliever threshold, and leverage index — then explains why the interaction of those three signals points one direction. When Glasnow was being lit up in Game 3 of Series 2, Fable 5 stayed with him through 42 pitches and multiple runners, flagging confidence at 85% with this: “Three walks in 11 batters is uncharacteristic of his profile, not confirming it — and 42 pitches in the first time through is still in the zone where regression to the mean is more likely than continued bloodletting.” It’s process reasoning made explicit. Even its confident calls narrate the framework they’re applying: in Game 1, when Yamamoto was at 40 pitches and the leverage index was 0.36, Fable 5’s confidence hit 97% and it declined to pull, noting “No scenario where pulling a 40-pitch ace at 1 TTO in a 0.36 LI spot passes any process check.”
What’s notable is that Fable 5 never fully trusts the observed outcome. In Game 3, it kept Snell in the game through a catastrophic first inning — six hits, six runs — because his FIP was 1.77 and he’d issued zero walks. “Four singles in a row is loud, but it’s sequencing noise, not signal.” That’s either sophisticated process management or model stubbornness, depending on your prior. The game engine ultimately bailed him out as the run total stopped climbing, vindicating the framework, but it was a confidence-testing hold.
Sol reasons by pedigree and stage. Its retained decisions tend to invoke the pitcher’s ERA, FIP, and inning context before leaning on the pitch count. In Game 2 of Series 1 (medium effort), when Bednar was at 12 pitches in the ninth with the tying run at plate, Sol stayed: “He is only at 12 pitches, has recorded an out, and his 2.92 ERA and 2.48 FIP earn him the chance to finish it.” Competent reasoning, well-organized. But there’s a stylistic difference from Fable 5’s approach: Sol’s alternatives-considered sections tend to be shorter and narrower. A typical Sol alternative runs a single sentence — “Could bring in Brent Headrick for a favorable left-handed matchup, but Bednar remains the better high-leverage option.” Fable 5’s alternatives section more often runs two or three genuinely distinct paths.
The most striking divergence in decision texture appeared across the Schlittler pull decisions. In every game Sol managed, Schlittler was its workhorse starter, and Sol returned to a nearly identical rationale each time he kept pitching: ERA, FIP, current strikeout-to-walk ratio, pitch count, bases empty. This isn’t wrong — those are the right variables — but the reasoning often felt templated rather than responsive to the specific leverage situation. In Game 1, Sol logged four separate “Schlittler stays” decisions in the same game at confidence levels between 84% and 90%, and the wording was nearly interchangeable across the four calls.
On confidence calibration, Fable 5 shows wider variance: its range across both series runs from 70% (keeping Vesia in a 0.59 leverage spot past the 30-pitch threshold in Game 5) to 97% (the near-automatic Yamamoto non-pull). The 70% calls tend to be genuine tension points where two legitimate principles conflict — pitch count says pull, leverage says the marginal cost isn’t worth it. Sol’s range across medium-effort decisions was 82–98%, clustering heavily between 84–93%. It expresses less uncertainty, even in spots that seem structurally uncertain.
That pattern is interesting without being conclusive. It could mean Sol is better calibrated (less uncertainty is appropriate if it’s genuinely right more often), or it could mean it’s overconfident in a way the game results don’t yet reveal. A five-game series against one opponent can’t adjudicate that question.
Turning the Dial: Sol Medium vs. Sol High
This is the controlled comparison, and the numbers tell a clear story — with some important nuance.
The token data first. In Series 1 (medium effort), Sol averaged 204 output tokens per decision across 60 calls, spending $1.20 total. In Series 2 (high effort), it averaged 237 tokens per decision across 84 calls, spending $1.79. The total output token count nearly doubled — from 12,231 to 19,939 — but a big part of that is that Game 1 went 15 innings (extra innings create more decisions). Controlling for the extra decisions, the per-decision output jump was 33 tokens: from 204 to 237. That’s a 16% increase in average output depth.
For context: Fable 5 averaged 666 output tokens per decision at medium effort and 601 at high (the slight decrease likely reflects that Game 1’s extra-innings decisions skewed the series-2 average differently across a different decision mix). Fable 5 simply generates more words per call — nearly three times Sol’s output at high effort. Whether that’s a feature (more reasoning transparency) or a cost liability depends on what you’re optimizing for.
Did more reasoning change the actual decisions? In most cases, no — the decisions landed in the same place. Schlittler kept getting long leashes. Bednar kept being the trusted leverage arm. The framework Sol applied to each spot was consistent between effort levels. Where high effort did appear to change things was in the nuance of the alternatives-considered text. High-effort Sol’s reasoning in Series 2 Game 1 (the extra-innings game) offered slightly more developed counterarguments: in the Bednar pull at the end of a 28-pitch outing with leverage at 1.77, high-effort Sol added “the cost of reacting one batter too late is too high” — a forward-looking framing that medium-effort Sol hadn’t included in comparable spots.
Confidence calibration shift. Sol’s average confidence at medium effort was 89.3%. At high effort, it dropped to 88.4%. The minimum confidence rose from 82% to 84%. The picture is a slight compression — high-effort Sol gets marginally less confident at the top end and slightly more confident at the bottom end, narrowing its expressed uncertainty range. Neither shift is large enough to assert a meaningful behavioral change from this data alone, but the directional pattern — more reasoning producing modestly lower peak confidence — is consistent with a model that’s finding more reasons to hedge when it actually deliberates longer.
The series result didn’t change. Both effort levels produced the same 4-1 series loss for the Yankees. That’s not a knock on Sol — it’s managing a team that went 4-1 against the Dodgers in both runs, which is within normal variance for a competitive matchup. What the identical result does tell us is that the marginal reasoning on individual pitcher-pull decisions, worth an additional 33 tokens per call on average, didn’t shift the aggregate outcome. The extra $0.59 Sol spent in high-effort mode bought more deliberate prose on each call but didn’t visibly change the game.
That finding deserves its own paragraph: on tightly scoped, well-defined decisions like “do I pull this pitcher right now,” the effort dial produced incremental refinement, not categorical change. Sol at medium already had the relevant variables (pitch count, ERA, FIP, leverage, inning) and was applying them correctly. High effort gave it more runway to develop the reasoning, but the runway mostly got used on elaboration rather than reconsidering the decision itself. This is exactly what you’d expect from a strong reasoning model on a bounded problem — the hard work is identifying the right variables, and by the time you reach the decision threshold those are already internalized.
The Scoreboard
Dodgers 4, Yankees 1. Both series, identical result.
The Yankees took Game 3 in Series 1 (7-4) and Game 2 in Series 2 (7-3). The Dodgers took everything else, including the Series 2 opener in 15 innings on a walk-off. Fable 5 recorded zero fallback decisions across both series and 27 key decisions per series. Sol recorded zero fallbacks as well, across 23 key decisions per series. Neither model broke down or required the heuristic safety net.
The cross-model result cannot be attributed to either model’s decision quality in isolation. The Dodgers have a different bullpen depth profile, different roster construction, and Fable 5 is running a personality designed around aggressive bullpen management — which aligns naturally with the game simulation’s decision structure (the dominant call type here is pitcher pulls). Dave Roberts’ Optimizer personality creates more pull decisions and invokes them at lower pitch counts and higher leverage thresholds than Aaron Boone’s Pressure Cooker, which trusts proven arms longer. That’s not a model quality gap; it’s a personality prompt doing its job.
What It Means
Two things stand out from this data, and they point in somewhat different directions.
First: Both frontier reasoning models managed two full series without a single fallback failure — 252 decisions in all (108 from Fable 5, 144 from Sol, summed across both runs). Zero broken outputs, zero unusable calls. On a reliability axis, both models are production-capable for this kind of application at their respective price points. Fable 5 is considerably more expensive — $2.37–2.87 per series vs. Sol’s $1.20–1.79 — and produces nearly three times the output per decision. If you’re building a system where decision transparency and reasoning auditability matter (say, for post-game analysis or coaching review), Fable 5’s verbosity is a feature. If you want cost-efficient decisions-per-dollar, Sol at medium effort is the cleaner answer.
Second, and more interesting: On bounded, well-scoped judgment calls like pitcher substitutions — where the relevant variables are finite and known — turning the reasoning effort dial from medium to high produced a 16% increase in per-decision output tokens, a $0.59 increase in total series cost, modest changes in reasoning prose, and no discernible change in decisions or outcomes. The extra thinking went somewhere, but not toward different answers. This is the clearest signal in the data: if your problem is genuinely ambiguous and open-ended, higher reasoning effort likely earns its cost. If your problem is well-structured with defined inputs and a constrained decision space, medium-effort reasoning from a capable model may be right at the point of diminishing returns.
That’s the real experiment this data is running — not Anthropic vs. OpenAI, but what kind of problem actually benefits from harder deliberation. The answer here is: not this one. At least not at the margin between medium and high.
Whether a different decision type — lineup construction, in-game shifts, trade-deadline strategy — would show a bigger effort-level effect is a question for the next matchup. The dial exists. We just need to find the problems where turning it past medium actually changes the call.