Can Claude Opus 4.8 Run the Lane? Opus vs Fable as Controller
What if the model you rely on goes away? We swapped only the controller and measured the difference.
Previously: Does Merging Beat the Best Solo Run?
The rounds so far have shown that results move depending on how the controller and the workers are paired. This time the strongest deliverable came from the Codex worker inside the Fable-led lane, with the Codex-led lane next. Which leaves an obvious question — what changes if you replace the controller itself with a different model?
At MIF we brought in Claude Code’s other model, Opus 4.8, as the controller, gave it the same task the Fable 5-led lane had run, and measured the difference. One thing to say up front: this was never an experiment designed to prove that Opus is better. It was built to observe what happens to the deliverables, the elapsed time, and the day-to-day operation when the same task and the same worker lineup are held fixed and only the controller changes.
This article covers one task with a single run per lane (n=1). It is not evidence about how these models compare in general. Some instruction conditions were not perfectly identical across lanes; where they differ, we say so in the text.
Why Opus — because Fable might have gone away
Why try Opus at all? The reason was practical. Fable 5, the model that had produced the best deliverables in every round so far, was reported to be at risk of dropping out of the flat-rate subscription plans. At the time of the experiment the situation read as “it may become unavailable after July 19.” If that happened, what would I move the controller to? The obvious candidate was Opus 4.8, since it is available inside the same Claude Code setup. So this round was not driven by curiosity — it was scouting a fallback.
The exact effective date and conditions for the subscription-plan change may shift. Regardless of what this article says, please confirm the current terms against official sources.
Update: the outcome was unknown while these experiments were running. According to Anthropic's official help center, from July 20, 2026 Fable 5 is included on Max plans and premium seats on Team plans, where up to 50% of the weekly usage limit can go to Fable 5 at no extra cost (this is not additive — other models draw from the same limit). On Pro plans and standard Team seats, Fable 5 runs on pay-as-you-go usage credits instead. Fable 5 also consumes those limits faster than other Claude models. So the premise that Fable would become entirely unavailable did not hold, depending on the plan. The experiments here were nonetheless run under the uncertainty that existed beforehand. Terms can change, so always confirm current conditions against official sources (as of July 20, 2026).
A further update: just before this article went out, Anthropic announced Claude Opus 5, the successor to Opus 4.8, on July 24, 2026. According to the official announcement, it approaches Fable 5's frontier intelligence at roughly half the price. Every experiment and score in this article used Opus 4.8. If you are choosing a fallback controller today, check the current model lineup and terms against official sources (as of July 27, 2026).
One other event pushed me toward it. According to Anthropic’s official support article, Fable has safeguards, and for certain content the model is designed to switch to Opus 4.8 automatically, even mid-turn (as documented when we checked in July 2026). While a fresh Fable session was reviewing the Round 3 deliverables, exactly that happened: partway through a turn, the model had automatically switched to Opus 4.8. Our fail-closed verifier — the check that requires an exact model match — caught it, and we invalidated that review. The safeguards behaved as specified and the verifier behaved as designed, which is good news on both counts. It also turned out to be an unplanned preview of Opus standing in for Fable. Having shown up unannounced, Opus was now going to be invited in as controller on purpose. The scouting trip was set.
Step 1: swapping only the merge step
I started the comparison small. Two finished candidates, the same evaluation summary, and the same merge instructions went to both Fable and Opus — a controlled replay that compares nothing but the merge judgment.
The result: the Opus merge scored 63.63, the Fable merge 47.84. A gap of +15.79. The Opus version made the conservative call, keeping the candidate with the safety design as the foundation and pulling in only the readable report display from the other one.
But this was a point to stop and think. That +15.79 covers only the step where two finished candidates get combined. It does not measure most of what a controller actually does: spinning up workers, getting two standalone implementations built, handling failures, and carrying the lane through to completion. The number was correct; the label on it was too broad. So I moved on to a full from-scratch comparison.
Step 2: handing the whole lane to Opus
Starting from the same public baseline that Fable and Codex had used, the Opus controller launched fresh workers and ran the full lane: two standalone implementations plus the Mix.
One point about the setup deserves to be explicit. In the Opus lane, the workers were still the same Fable and Codex as in every other lane. An “Opus controller with Opus workers” configuration was never tested. I worked on the assumption that Fable is the better Claude-family worker — an assumption, not something I verified — and changed the controller alone. Swapping the workers too would have made it impossible to tell whether any difference came from the controller or from the workers.
The run itself was not smooth.
- Asked — under the pre-registered protocol and execution harness, the Opus controller started the lane
- Stalled — the Windows sandbox configuration failed to come up and had to be fixed. Then, mid-run, provider capacity on the worker side (a temporary supply-side limit) interrupted the lane twice
- Came back — an auto-resume mechanism scoped strictly to capacity-related failures brought it back, and with human approvals along the way the lane ran to completion
- Measured — all 9 deliverables (3 lanes × 3 outputs) were frozen and scored together on identical data generated from a fresh seed (Eval4)
Step 3: same workers, different controller — here are the numbers
Eval4 lines the lanes up side by side. The thing to watch is what score the same kind of worker produced under each controller.
| Deliverable type | Fable-controlled lane | Opus-controlled lane | Gap |
|---|---|---|---|
| Codex solo | 65.07 | 55.89 | Fable side higher by 9.18 |
| Mix | 47.84 | 48.15 | Opus side higher by 0.31 |
| Claude/Fable solo | 44.30 | 42.92 | Fable side higher by 1.38 |
The largest gap is the Codex solo at 9.18 points. The same Codex worker scored higher when it was working under the Fable controller. The Mix, meanwhile, came out 0.31 ahead on the Opus side — consistent in direction with the merge-only replay, where Opus made the better call. But the top of the whole field (65.07) stayed out of reach: the best score in the Opus lane (55.89) fell 9.18 points short of it.
I am not calling that gap a difference in raw model ability. The fine detail of the worker instructions and the execution history were not fully aligned across lanes, and each is a single observation. What can be said is a fact about the two measurements: the merge-only comparison (+15.79) and the full-lane comparison (9.18 in Fable’s favor) point in opposite directions. Measure a part or measure the whole, and the verdict on the controller flips. That was the biggest finding of this round.
Where the hypothesis stands — how to pair controller and workers
Based on what I saw, this is my current working hypothesis.
- The most promising setup is a Fable controller with a Codex worker. The top score across all nine deliverables (65.07) came from that pairing.
- If the controller moves to Opus, the Codex worker still looks like the backbone. The best score in the Opus lane also came from the Codex worker (55.89). It did not reach what the Fable-led lane produced, but Opus did demonstrate that it can carry a lane to completion as controller.
- Or move the controller to Codex as well. In the Codex-controlled lane the Codex-worker deliverables scored 58.32 (Mix) and 57.07 (solo), both above the best result from the Opus lane this time.
All of these are hypotheses drawn from one task and one run each, not conclusions. Still, the common thread — under every controller, the Codex worker was the backbone of the result — is a practical hint when you are shopping for a fallback. And since the merge-only comparison (+15.79) and the full-lane comparison reversed the verdict on the controller, there is a second lesson to hold onto: before asking “which model makes a better controller,” check which stage of the work you are actually measuring.
Next time: the finale
Changing the controller did not change which deliverable came out on top. But the experiment left behind a question that no model swap can answer. What if a specific model becomes unavailable? What if a stronger one ships tomorrow? The final installment looks back at how this three-way contest ended and connects it to our answer — ModelOrcs, a collaborative development engine built around a setup where both the controller and the workers can be replaced.
This series (6 parts)
- Same Question, Different Proposals
- 100 Connect Four Games Head-to-Head
- Claude Code vs Codex on Real EC Data
- Does Merging Beat the Best Solo Run?
- Can Claude Opus Run the Lane? Opus vs Fable as Controller (this article)
- Three Controllers Compared, and ModelOrcs