The Engineer

Every launch-week price comparison hinges on one variable the pricing page leaves out: the cache-hit rate. AINews ran the Artificial Analysis data. Muse Wonder burns about 7.69M input tokens and roughly 54k output tokens per benchmark task. Output is about $1.08 of a $4.52 bill. The rest only reconciles if around 93% of input tokens bill at the $0.20 cached rate instead of $4. Drop the hit rate to 80% and the task costs about $8.46. At 50% it is about $17, more than double Fable 5.1's $7.63.

The curve is steep because agent loops are input-bound. Each turn re-sends the system prompt and tool schemas along with all accumulated context. Artificial Analysis now publishes measured cache-hit rates and per-provider cost per task. The spread is wide. GLM-5.3 Max ranges from $0.92 to $8.68 per task across providers, and most of that gap is cached-input pricing.

The Nov 30 flip

Gemini 4 Argon's $2/$10 rate, with a 95% cached-input discount, is introductory. AINews assumed flat token use and every rate component doubling. Under that assumption Argon moves from $1.99 to about $3.98 per task. That is roughly 20% above GPT-6 Astra's $3.26, if Astra's price holds. Devin's 75% Argon discount ends sooner, on Oct 27.

Where the cheap models break

Price is irrelevant if the output is wrong. Practitioner Theo reports Argon scored perfectly on his Skatebench-private test and made false claims on over half of his codebase-analysis questions. Ethan Mollick reports overconfidence and weaker execution on complex multi-agent work. Arena ranks Argon High fifth on its agent leaderboard. Two of Argon's coding wins, 52.8% on CursorBench and 60.5% on FrontierCode 1.1, come from Cursor and Cognition. Both companies sell Argon access.

Wonder's evidence is thinner. Techpresso notes Meta's own model card calls its competitor runs 'best-effort'. The one hands-on review, from ben's bites, found Wonder slow and verbose, and poor at UI work. Verbosity is a cost term. Wonder is 60% cheaper per token than Fable, so the advantage is gone once it uses more than 2.5x the tokens per finished task. Wonder Full, the stronger variant, ships only inside Muse Code with no token price. Its 68 on the Coding Agents Index scores the model plus Meta's harness. Nobody running a different loop can reproduce it.

The routing rule

Route by whether the output is checkable. If a compiler, test suite or schema checks it, send it to the cheaper model and let the check catch failures. If the output is a claim a human will trust (code review, root-cause analysis, document Q&A), it needs evidence anchors and a second pass. AINews flags a cheap pattern. The LongHarness authors report GLM-5.3-Flash going from 4.4% to 61.1% exact match on OfficeQA Pro V2 just by holding evidence and conclusions in separate, explicit state. Nobody has tested this on Argon yet. It is still a cheaper first experiment than another model swap.

A per-token price is a quote. Cost per resolved task, at the measured cache-hit rate and the post-promo rate, is the bill.

A gate that passed 46 tests and still failed open

Daily Dose of Data Science took apart an incident-response agent that Claude Code built from a single prompt: 30 files, 3,383 lines and 46 passing tests. A decision model returned a chosen action plus a probability for each of six options. A deterministic gate then ran the action if its probability was 0.80 or above, gathered more evidence between 0.55 and 0.79, and handed off to a human below that. The gate itself was correct. The bug was in the parser in front of it. When the chosen action was missing from the probability map, the code used the largest probability of any option instead. So a response that picked page_team, with only inspect_logs listed at 0.90, became page_team at 0.90, which clears the execute threshold.

No test caught it, because every fixture and every live response included the chosen option's probability. The broken rule is that an action must stay paired with its own probability, and that rule spans the parser, the gate and the dispatcher, so no single file owns it. A review by CodeRabbit, the post's sponsor, found the bug along with four others. Incident text was rendered as stored XSS. Shared records had no authorization checks. The model-endpoint URL was unrestricted. Fixture mode fell through to live requests.

The same flaw in three more places

AINews also reports a study of 11 multimodal models in which giving them tool access raised harmful-request refusal failures for every model, by up to 68.7%. Safety evals run in chat mode tell you little about how your tool-using agent behaves.

Why prompt rules lose

CSO Online doesn't describe the echo mechanism, so this is our reading. Control and data travel in the same channel. The framework parses whatever the model emits and runs it. Asking a model to repeat a string looks harmless, so a system-prompt rule like 'never repeat tool calls' is a probabilistic defense against output the model was explicitly asked to produce. Six frameworks failing to one prompt points to a shared design assumption rather than six identical coding mistakes. The fix looks like the fix for SQL injection: structural separation plus authorization in code the model can't influence. The MCP findings make the risk worse. Injection decides which call fires, and a server averaging about two verified bugs decides what that call can do.

The smart move

Give each agent one place where model output is parsed. Look up confidence by exact key. Return zero or reject the response when a key is missing, a value is NaN or out of range, or an action falls outside the allowed set. Make the gate the only component that can issue an executable-action token, and have the dispatcher refuse any call without one. Then attack your own stack. Plant a well-formed tool call in an email body, a RAG chunk or a ticket, and ask the agent to repeat it. That test takes a few days, and it will tell you more than the unnamed framework list will.

If a missing or echoed value can reach your executor, your gate isn't protecting anything, whatever your green tests say.

Newman describes disaster recovery as four patterns. Cost rises as recovery time falls. One system can mix patterns across its parts. The capacity risk sits in the two cheapest, because both acquire compute after the failure.

The Pragmatic Engineer's excerpt has Newman's clients struggling to secure capacity. Moving regions only sometimes helps. Server lead times keep growing. Some services can be reserved and some cannot. AWS sells guaranteed EC2 availability, billed whether it runs or not. Guaranteed Lambda invocations are not for sale. When a region fails, every tenant in it goes looking for the same spare capacity at the same moment. Newman offers three options. Reserve capacity. Prioritize critical services. Or run fully redundant infrastructure and give it other work. In his BankCo story, idle DR desktops ran as extra calculation workers until a disaster needed them. The modern equivalent is a ring-fenced load-test environment that gets aborted when DR kicks in. That is good engineering.

The model provider is the weaker dependency

Newman says foundation-model providers have weaker availability records than cloud vendors, so a hosted LLM needs its own failure plan. His backup routes are OpenAI on Azure and Anthropic via Google's Gemini Enterprise Agent Platform. The backup may lag on the newest model. Prompts tuned for one model do not fully carry over to another. A failover route is real only if the evals show the feature still works on it. He also recommends timeouts that scale with tokens generated, plus monitoring of when streaming starts. A single latency SLO over whole responses misleads when users write the prompts. Each vendor added to one request flow multiplies its own availability into the total, so each hop needs its own fallback.

Other reporting supports the scarcity premise. In Fortune's Term Sheet, investor Hans Tung argues: 'The models that win are the ones that are available.' He was an early Anthropic backer, so a compute-scarcity story suits his portfolio. The Information reports Anthropic has signed $531B of compute deals across AWS, Google and Microsoft in 11 months. The labs behind those APIs rent from the same hyperscalers a failover would land on. Microsoft's Maia 300 does not fix this soon. Microsoft has built only tens of thousands of custom chips. 'Multi-gigawatt' deployment implies millions.

Where the evidence is thin

Newman's capacity evidence is anecdotal client conversation. Gergely Orosz's claim that AI-driven deploy volume is causing more outages has no data behind it. The Coinbase example is the useful check. A single AWS zone problem took Coinbase's global trading service offline. A system that cannot survive a zone failure automatically is a long way from regional DR.

Newman also argues RPO and RTO make poor SLOs. Real disasters are too rare to produce the measurements an SLO needs, so a drill is the only evidence available. The cheapest drill is a restore to a secondary site inside the next load test, using the same Terraform as production.

A DR plan that provisions capacity at failover time bets that scarcity goes away. An LLM vendor with no tested fallback is the same bet.