Which models can run Claude Code through DIAL?Five upstreams, and the cost winner is open-source.

Does the harness route, and does it survive the swap.

Before you can swap in a cheaper model, two things have to be true, and only one is plumbing: Claude Code has to reach the model through the gateway at all, and the harness has to keep working when you point it at an open-source upstream instead of an Anthropic one.

The first question is settled quickly; the second is the one that decides your bill. Its answer is that the harness is portable at Sonnet scale, not at Haiku scale, and the determining variable is model scale, not vendor.

Sustain a plan, not just emit a tool call.

"Does open-source work with Claude Code?" collapses on contact, because most modern coder models can emit a single tool-use block. The harness needs more: the model must plan, observe a tool result, update its plan, and continue. The gating question is agentic-loop sustainment, not structured-output emission. Two falsifiable hypotheses frame it.

• H1, plumbing. Claude Code routes through the gateway via an Anthropic-to-OpenAI shim with no harness changes. Falsified by any partial-message rate above zero across 100 consecutive runs, or any tool-use round-trip divergence between the local-Anthropic and gateway-Anthropic baselines.
• H2, portability. Harness viability with non-Anthropic upstreams is a function of model scale; sub-100B active-parameter models cannot sustain the multi-step loop. Falsified by a sub-100B open model passing the multi-step tasks at a rate comparable to Haiku 4.5.

One adapter, two routes, one wire surface.

The bridge is the only custom code between Claude Code and the upstream. It exposes Anthropic's messages shape on the wire, translates each request into the gateway's OpenAI-shaped chat-completions API, and streams the response back in the shape the harness expects. Two wrapper configurations differ only in which upstream the harness's three model slots point at; everything else, the agentic loop, the system prompt, the tool inventory, is untouched. One behaviour is load-bearing: the bridge strips cache-control markers when the upstream is not Anthropic, because the open routes otherwise return a no-route error.

Eight tasks, escalating so the shape of failure carries the diagnosis.

Each task is self-contained Python under a hundred lines, with a hidden validator that exits non-zero on any deviation, removing LLM-judge bias. The portfolio climbs from one-shot answers to the full agentic loop, where the hardest task requires the model to remember what it just did across a tool call and update an external surface from memory. No model lands between "passes all multi-step" and "passes none": failure is at the loop level, not the task level.

Pass rate misleads without cost per success.

A 38 percent pass rate could mean "half competent" or "passes the easy tasks and bricks the hard ones in a structured way," so the metric portfolio is designed to disambiguate. Hidden validators provide outcome ground truth without an LLM judge, whose sharpest failure in the prior bake-off was over-rating models whose prose style resembled its own.

The cheapest model that passes everything is the big open one.

At Bedrock-verified pricing, the open 480B model is not just viable, it is the floor. The earlier bake-off had it looking expensive only because the SDK billed unrecognised models at the Opus fallback rate; the console rates tell the real story.

One honest uncertainty: the Anthropic 4.x Bedrock rates come from the SDK's internal table, which agreed with verified rows within about five percent but was not independently console-checked. And the gateway contract price is the only number that matters in production; the ranking survives any contract that does not inflate the 480B input rate above roughly a 6.7x markup.

Capability tracks parameter scale, not vendor.

The open 480B model's input-token total exceeds Sonnet's on the same workload for a structural reason, not a model one: the gateway forwards cache-control intact for Anthropic, so Bedrock caches the 20K-token system prompt and reads it back at about a tenth of the cost, whereas the bridge must strip cache-control for the open routes, so they pay full rate every turn. That is a gateway gap, not a model gap; cache support for non-Anthropic routes would narrow the cost difference an estimated 30 to 50 percent. On latency, per-call the open 480B is actually slightly faster than Anthropic; the long tail comes from turn count and tool-result round-trips, so the harness makes the tail.

The closing thesis: harness viability is scale-driven, not vendor-driven. Three open lineages tested across this and the prior bake-off, a 27B, a 123B, and the 30B here, all fail with the same one-call-then-stall signature; only the 480B, roughly an order of magnitude larger in active parameters, closes the loop. Open models keep up at roughly 400B-plus active parameters on a coder-tuned base, and not below.

Put the 480B in the Sonnet slot. Never the 30B in the Haiku slot.

• For operators. Adopt the open 480B as the Sonnet-tier default; never put a sub-100B open model in the Haiku slot. The economics and the loop both favour it.
• For gateway operators. Ship cache-header support for non-Anthropic routes; it is the single change that most narrows the cost gap.
• For open-model trainers. The target is multi-turn tool-use transcripts with sustained plans, not single-shot tool-call benchmarks.
• For replication. The real cutoff lives somewhere in the 30B-to-480B gap; test 70B to 150B coder models to find it.