Claude Code + Codex: Does Merging Beat the Best Solo Run?

Previously: Claude Code vs Codex on Real EC Data

“If you have two good builds, mixing the best parts of each should give you something better.” Intuitively, most people would agree. But merging two finished products is not simply a matter of adding strengths together. Adopt one side’s design and you can break an assumption the other side was relying on.

We at MIF have been running an experiment in which Claude Code (model: Fable 5) and Codex (model: GPT-5.6-Sol, reasoning effort: ultra) are given the same task and set against each other. Part two was a Connect Four bot; part three was a business tool for finding sales opportunities in e-commerce data. This time we take up the question that follows from those: can a “Mix” that fuses the work of two AIs beat either one alone?

The short answer: mostly no — but exactly once, marginally yes. And the way those cases split is where the practical lesson lies.

How a “Mix” of AI outputs gets built

Start with the process. A Mix is not two artifacts mechanically bolted together, nor is it two AIs co-writing while consulting each other. It goes through four stages:

1. Complete two solo builds — a Claude worker and a Codex worker each take the same task all the way to completion, without seeing each other’s work in progress
2. The orchestrator compares them — the AI running the lane (the orchestrator) reads both finished builds side by side and analyzes their strengths and weaknesses
3. Merge into a third artifact — it decides which build to use as the base, what to port over, and what to protect, then implements the Mix
4. Re-evaluate on the same yardstick — the Mix is scored by the same referee and the same criteria as the solos, then compared against them

The experiment runs two lanes: Claude orchestrates Lane A, Codex orchestrates Lane B. Each lane produced two solos and one Mix, for six artifacts in total.

The thing to keep in mind is that a Mix is itself just another proposed change. Adding features can still leave the whole weaker if the consistency or the safeguards the original build maintained fall apart. The value of a merge isn’t determined by spotting strengths alone. You can only judge it once you have decided what must not break and re-measured on the same yardstick.

Round 1: even the winner of the final fell short of the strongest solo

Round 1’s Connect Four contest was settled not by human scoring but by match records on a shared referee program. To rule out convenient tuning after implementation, the bots played from held-out opening positions, swapping who moved first.

The result was clear. In neither lane did the Mix come out ahead of its own lane’s strongest solo. Fusing two roughly equal engines did not raise the level of play.

In the final between the two lanes’ Mixes, meanwhile, the Claude/Fable-led Mix won 57, the Codex-led Mix won 35, with 8 draws — a wide margin.

These two results don’t contradict each other. The final tells you which of the two Mixes is stronger; the within-lane comparison tells you whether merging produced something stronger than the best solo. Even the winner of the final wasn’t necessarily improved by the merge — in Round 1, the Mix’s win/loss record and the merge’s effectiveness had to be read separately.

In a game bot, the two solos are chasing the same winning lines on the same board. Stacking similar capabilities produces competition more readily than complementarity, and what stood out was the downside: the merge process breaking consistency instead. That said, this is a hypothesis formed from these results, not a proven cause.

Round 2: on a business tool, the outcomes split

Round 2’s task was a tool that proposes, with supporting evidence, the actions worth taking this week based on e-commerce inventory, order, and customer data. Unlike a game with a single winning line, it has to handle detection coverage, data quality, safety, and how actionable the proposals are all at once — which left room to transplant strengths across different designs.

We froze all six artifacts and scored them on newly generated held-out data (Eval 3), and an interesting split appeared.

• The Claude/Fable-led Mix came in 17.29 points below the best solo in its own lane on Eval 3. It improved on one of its two bases, but came nowhere near the lane’s strongest solo.
• The Codex-led Mix came in 1.15 points above the best solo in its own lane on Eval 3. Across both rounds, this is the only case where a Mix beat its lane’s strongest solo.

As a check, we recomputed the same gap in a follow-up evaluation (Eval 4) that generated fresh held-out data. The Codex-led Mix’s margin there was 1.25 points; the direction didn’t change. But Eval 3 and Eval 4 are separate yardsticks. We haven’t computed any differences across the two, and because Eval 4 re-scores the same artifacts, it doesn’t add an independent task either.

Looking inside, the Codex-led Mix picked up a detection type that neither solo in its lane had caught (cross-sell proposals) during the merge, while losing ground on data-quality scoring — netting out to a small positive. And although it became the best in its lane, the top candidate overall was a solo from the other lane. Merging alone did not put it at the top.

The Fable-led Mix’s heavy loss needs a footnote too. Part of the gap comes from factors outside the merge itself, such as deductions for submission format and structural penalties in the safety scoring. Even accounting for those, though, the conclusion that it fell short of the best solo doesn’t change.

And merging isn’t free. It adds a whole stage of comparison, transplanting, and re-verification. In the Codex-orchestrated lane, where we could measure usage, the Mix took several times the wall-clock time and tokens of that lane’s solos (the Claude-orchestrated lane was not measured). What that bought was, at most, 1.15 points on Eval 3.

Mix vs. solo, side by side (Eval 3 and Eval 4)

Round	Evaluation	Lane (orchestrator)	Strongest solo vs. Mix
Round 1	Referee matches (held-out positions)	A (Claude)	Mix did not beat the strongest solo
Round 1	Referee matches (same)	B (Codex)	Mix did not beat the strongest solo
Round 1 final	Referee matches (Mix vs. Mix)	A vs. B	Claude/Fable-led Mix 57 – 35 Codex-led Mix (8 draws)
Round 2	Eval 3	A (Claude)	Mix trails the best solo by 17.29 points
Round 2	Eval 3	B (Codex)	Mix leads the best solo by 1.15 points
Round 2	Eval 4 (re-scored, new seed)	B (Codex)	Mix leads the best solo by 1.25 points (same direction as Eval 3)

Does mixing Claude and Codex make it stronger? The one condition that held

Across both rounds, one hypothesis emerged.

Merging works when the two builds’ capabilities are complementary. The Codex-led Mix gained its small margin in Round 2 because the merge picked up a detection type (cross-sell proposals) that neither solo in the lane had caught. When the missing piece is clearly identified, merging has a point.

Conversely, mixing roughly equal capabilities does not add up to more strength. In Round 1, both engines were already strong enough that there was almost nothing new to gain from combining them — leaving only the risk of breaking consistency in the merge.

And there is one more fact worth remembering. In Round 2 (Eval 3), the top of all six artifacts wasn’t a Mix at all. The highest score went to a solo built by a Codex worker under a Fable orchestrator (second place went to an artifact from the Codex-orchestrated lane). Before getting clever about mixing to make something stronger, finding a well-matched orchestrator/worker pairing mattered far more in what we observed — that is the biggest message in the ranking table.

The other lesson is that the value of a merge should be measured by the conditions it preserved, not the features it added. A vague instruction to “take the best of both” won’t stop a feature addition from regressing another feature, data quality, or safety.

If you’re building a Mix in practice, the procedure looks like this:

• Take the solo that scored highest in evaluation as the base
• Spell out, one by one, the gaps you want to fill
• Set regression conditions on the detections and safety properties that must be maintained
• Re-measure with the same evaluation once it’s done

In other words, treat it as change management, not as adding up ideas. That was the reliable approach.

Differences in conditions we should disclose

There is an asymmetry in this comparison that we should state honestly.

In Round 2’s Lane A (Fable-led), the orchestrator saw the identity of each candidate — which AI wrote which — plus a summary of the weaknesses flagged in an earlier evaluation, before merging. In other words, it was not blind. In Lane B (Codex-led), by contrast, the candidates were anonymized and no earlier evaluation was handed over.

On the face of it, the side with more information should have had the advantage. Yet the one that beat its best solo was the side working with anonymized candidates. That still doesn’t mean prior information is harmful, or that going blind is better. Anonymization can’t fully erase identifying cues, and the Codex orchestrator did see a summary of the Claude side’s results before its lane began. The instructions given to the workers weren’t identical across lanes either.

So this is not a controlled experiment in which only the merging agent was swapped. You cannot read the score gap between the two Mixes as a difference in orchestrator capability.

The evaluation has its own boundaries. What Eval 3 measured was not just the application itself but the full submission, including the execution and formatting components — and implementations that fit the scorer’s classification scheme can be favored. Point gaps are an important observation, but they don’t capture all of a product’s value in a single number.

We have one task per round, two cases in total. We won’t generalize to “Mix is pointless” or “Mix works.” In what we observed, merging fell short of the strongest solo in almost every case, and produced a small gain only in the one instance where it filled a capability gap. That is as far as it goes.

Finally, the standings, once more for the record. Of the six artifacts in Round 2 (Eval 3), the highest-scoring was a solo built by the Codex worker in the Fable-orchestrated lane. Next came the artifacts from the Codex-orchestrated lane (the Codex-led Mix, then the Codex solo). Same worker, different result depending on which orchestrator it worked under — that observation is what led to the next experiment.

Next time: what changes when the orchestrator becomes Opus?

The Fable-led Mix sank badly, and only the Codex-led Mix floated slightly. So what happens if we swap out not the worker but the orchestrator itself for a different model, and give it the same task?

There is a pressing reason behind that question. Fable 5, which has produced the best artifacts so far, was set to drop off the flat-rate plan — at the time of the experiment, it was reported that it “might become unavailable after July 19.” If that happened, what do we replace the orchestrator with? Next time we put Claude Code’s other model, Opus 4.8, in the orchestrator’s seat and measure the gap against the Fable-led lane.

This series (6 parts)

1. Same Question, Different Proposals
2. 100 Connect Four Games Head-to-Head
3. Claude Code vs Codex on Real EC Data
4. Does Merging Beat the Best Solo Run? (this article)
5. Can Claude Opus Run the Lane? Opus vs Fable as Controller
6. Three Controllers Compared, and ModelOrcs