Field notes on AI coding agents: what they cost, where they break, and what shipped. Figures without a traceable source were cut.
Written with AI assistance. Figures without a traceable source were cut before publishing.
OmniRoute aggregates 237 providers and advertises roughly 1.6 billion free tokens a month, and that figure is arithmetic rather than a promotion, because the RTK+Caveman layer compresses 10,000 tokens down to 1,080, and a free quota multiplied by about 10 is exactly what a ratio like that buys you. What the number does not tell you is which model answers your next request. That is the only thing your bill and your output quality actually depend on.
The routing layer sits in front of 24+ coding tools and exposes a single endpoint, so the configuration work that normally multiplies across every tool in your stack collapses into one place. Compression does the rest.
Ten thousand tokens becoming 1,080 is a claim you can check against your own logs, which is more than can be said for most efficiency numbers, and checking it takes an afternoon rather than a procurement cycle, because the whole test is running one representative prompt through the router and reading the token counter on both sides of it. Do that before you rearrange your stack around the quota. A compression ratio measured on someone else’s workload is a hypothesis about yours.
Underneath that endpoint is a 4-tier fallback, and your subscription is the first rung.
When the subscription is spent the router moves to paid API calls, when the API budget runs out it drops to cheaper models, and when those run dry it lands on free-tier options, which is how the vendor can honestly say your coding never stops. Read the sequence again as a cost policy rather than as an uptime feature. Uptime is what gets advertised. What actually gets decided is which model you are talking to on the twenty-eighth of the month, and I have not seen anyone publish what output looks like at the bottom of that ladder.
The failure mode here is not an outage. It is a quiet substitution.
The useful thing about the current OpenAI lineup is that the tiers are documented well enough to reason about, which makes the fallback question concrete instead of philosophical. Sol is the heavy one, with a claimed 54% better token efficiency than models at the same level and an Ultra mode that runs 4 sub-agents in parallel. Terra extracts complete data sets and produces standard analysis documents. Luna answers fastest, hits 95% accuracy on basic question answering, and costs the least per batch.
Those are three different products wearing one brand. A fallback from Sol to Luna is not a small degradation in a continuous quality dial, it is a switch from a model that plans multi-step work to a model tuned for high-concurrency support tickets, and if that switch happens silently at 2am your build output changes character without anything in your logs saying why.
Cost differences at the top end are large enough to make the point on their own. One front-end benchmark put GPT-5.6 Sol at $1.43 per run against $9.05 for Fable 5. A 6x spread on a single evaluation is the kind of gap that should decide your default long before a free quota does.
Time behaves the same way. Sol’s Ultra mode took 2 hours 41 minutes 35 seconds to build one detailed 3D scene, running its sub-agents in parallel the whole time, and a run of that length is not something a fallback can politely interrupt halfway through. The reported side effects were heavy token consumption and a browser that struggled to render the result, which is a reminder that the expensive part of a long agent run is rarely the part you budgeted for.
If the fallback changes your output, the honest response is to measure the output rather than the uptime. Published agent evaluation practice splits into five dimensions: functional correctness, process quality, efficiency and cost, robustness and security, and experience and alignment. Correctness is treated as P0 and expected to be fully automated. Process quality is P1 and combines rubric scoring with human spot checks. Efficiency and cost is P2, and the method is baseline comparison plus token consumption monitoring.
That third one is the whole game for anyone routing through an aggregator. Baseline comparison plus token monitoring is precisely the instrumentation that tells you a fallback fired and cost you something, and it is also the instrumentation almost nobody sets up before the bill arrives.
The same body of practice recommends three kinds of judge working together. Deterministic scorers verify hard facts such as whether a tool was called or a file exists. Rubric scorers handle structured natural-language output. Human scorers are reserved for high-stakes domains like medical and financial work. A router that silently swaps models is a strong argument for owning at least the deterministic layer yourself.
There is a second audience for all of this that changes the math, and it is not developers. Sales staff using ChatGPT Work reportedly turned a CRM export into a technical proposal in 24 hours, work that had previously taken weeks, with something on the order of 70% of the time removed. One projection has non-coding users going from 20% of that product’s base to 60% within 12 months. If most of the requests hitting your routing layer come from people who cannot tell a good answer from a fluent one, the deterministic scorer stops being a nice-to-have.
Look at what tokens cost when nobody is giving them away. Top-tier Chinese models such as GLM5.2 and DeepSeek V4 Pro sit near $1 per million tokens at inference gross margins of 10% to 20%. At the low end, MiniMax M3 runs $0.06 to $0.2 per million and draws 60% to 70% of its revenue from outside its home market. DeepSeek introduced peak and off-peak pricing in mid-July, with the peak rate at twice the off-peak rate.
A spread from $0.06 to $1 per million is more than 16x, and peak pricing adds another factor of 2 on top, which means the decision of which tier to route to on a Tuesday afternoon moves your bill further than any free quota can. Meta priced Muse Spark 1.1 at $1.25 per million input and $4.25 per million output, roughly 75% and 83% below Anthropic’s Opus, and the tradeoff is visible in the benchmarks, since it leads on MCP Atlas and JobBench while trailing on SWE-Bench Pro and DeepSWE 1.1.
Specialization beats aggregation on cost more often than the aggregator story suggests. Alibaba’s Open Code Review, tested across 200 real pull requests and 50 open-source repositories covering 10 languages, scored higher on accuracy and F1 than a general-purpose agent while consuming roughly one ninth of the tokens.
Nothing in a routing layer gets you a 9x reduction. Picking a tool built for the job does.
The architectural advice worth taking from all of this is older than any of these products. Build the model-calling layer so it can be replaced, keep planning in the cloud and execution local where that split makes sense, and treat every aggregator as something you should be able to walk away from in an afternoon.
Tooling for that already exists in the ecosystem. Codex++ injects multiple API gateways into a Codex configuration in one click. CC-Switch reaches DeepSeek and GLM without a ChatGPT account at all. MonkeyCode hands out 900 million free tokens with no aggregator in the middle. None of those are better than OmniRoute in general, and that is the point, because a routing layer earns its place by being easy to leave rather than by being the only door.
I could be wrong about how much the fallback actually degrades things, and the honest answer is that neither I nor the vendor has published a number for it. What I would do before trusting a quota that large is run the same three prompts against each tier, record the token count and the wall-clock time, and keep the results somewhere I will find them the week the subscription runs out. The 1.6 billion is real. It is just not the number that decides anything.
Also readable on Telegraph.
Read next
Part of ai-coding-field-notes — field notes on AI coding agents. Found something wrong, or shipped something similar? Open an issue — corrections are the point.