Theo's “Meta just caught up with Anthropic” makes a good starting point for this week's model conversation. Muse Wonder puts another serious option in front of developers. Sonnet and Haiku 5.5 make the same conversation interesting from a different direction: how much capability do we need for each piece of work, and what does it cost to keep moving?
For Codex users tired of the reset game, that last question has some weight. A more capable model is welcome. A working day that depends less on refreshing a usage meter is welcome too.
Announcements, pricing and public discussions checked October 13, 2026. These are published claims and reported experiences; I have not run a comparative model benchmark.
Theo's enthusiasm comes with useful qualifications
The video is chiefly about Muse Wonder Full. Its best argument is more specific than a declaration that Meta has won.
Around 11:14, Theo moves from disappointing interface demos to an audit response he finds unusually readable. Around 14:13, he discusses paired coding work: Opus produced deeper, more correct work in the example, while Muse contributed speed and useful review. Near 36:19, he explicitly discourages dropping a Claude Code subscription for it before the value of Meta's allowance is understood.
Those are his observations on his tasks. They make me interested in trying the model on audits and bounded changes, while keeping the acceptance check in my own hands. Readability matters when I need to understand a finding before approving a fix. So does whether the fix survives a build and a runtime check.
Muse Wonder and Wonder Full are different buying decisions
Meta launched both models on October 12. The company emphasizes memory, sustained work and coding: retaining relevant context, working through blockers, and continuing through testing and revision. These are useful ambitions for an agent working on an actual repository.
The access distinction matters immediately. Meta's model page lists Muse Wonder through its API and Muse Code, with a 1M-token context window and rates of $4 input / $20 output per million tokens, plus $0.20 cached input. Muse Wonder Full also lists 1M context, but access is through the Muse Code Power Usage and Maximum Usage subscriptions. The page does not publish a per-token API price for Full.
That means the standard model's price cannot be attached to Full's results. Nor can a subscription's advertised usage multiplier tell us how many of our own projects it will finish.
I would start a trial with something familiar: a bug I can reproduce, an audit with findings I can inspect, or a change with a clear completion condition. The interesting question is whether Muse adds a useful second opinion or finishes work at an acceptable total cost. Another provider creates an option; the task still has to establish its value.
Sonnet 5.5 makes the everyday model more interesting
Anthropic introduced Sonnet 5.5 on September 28, positioning it for well-scoped work, bug fixes and polished documents. Its launch claims include 30%+ faster output generation than Sonnet 5 and up to 30% lower cost per task in Anthropic's testing. Those describe different measurements: generation speed and the cost of completing a workload.
The input and output prices stayed at $2 / $10 per million tokens. On October 7, the Haiku launch also cut Sonnet's cache-read rate from $0.20 to $0.10 per million tokens. Anthropic estimates that change makes most Sonnet agentic work around 20% cheaper. That is a vendor estimate for workloads with substantial cache reuse, rather than a discount every individual request will receive.
This is the part I would examine in an existing coding workflow. Repeated repository context can make caching consequential. If the model also needs fewer steps to finish a bounded task, the benefit compounds. If it creates another hour of review, the nominal saving matters less.
Haiku 5.5 makes smaller jobs worth separating
Haiku 5.5 arrived October 7, aimed at quick, repetitive and cost-sensitive work: summaries, classification, compaction and narrowly scoped subagent tasks. It adds an adjustable effort setting. Anthropic still points complex agentic coding toward Sonnet and Opus.
The pricing has a threshold that deserves more attention than the headline. For prompts up to 100,000 tokens, Haiku costs $0.10 input / $0.50 output per million tokens. For prompts over 100,000 tokens, the rates become $0.50 / $2.50. Cache reads also rise from $0.01 to $0.05. Check those current terms before budgeting a workflow.
That makes the size of a handoff part of its economics. A small worker locating a symbol, extracting facts from a document, or summarizing a test failure can be useful. Sending it the whole repository by default may erase part of the saving and make the assignment harder to follow.
I would give it a narrow question, the necessary context and an output I can check. Escalation belongs in that workflow too: sometimes the cheaper first attempt is just an extra attempt.
Put the price beside the model you can actually use
These are listed US-dollar API rates per million tokens, checked October 13. They exclude cache writes, tools, retries and provider-specific adjustments. Full's subscription access is shown separately because it has no listed API token rate.
| Model / tier | Input | Output | Cache read |
|---|---|---|---|
| Muse Wonder | $4.00 | $20.00 | $0.20 |
| Muse Wonder Full | Subscription | — | — |
| Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
| Haiku 5.5 ≤100K | $0.10 | $0.50 | $0.01 |
| Haiku 5.5 >100K | $0.50 | $2.50 | $0.05 |
The Haiku tiers refer to prompt length. Dashes indicate that Full has no listed API token rate.
On the smaller Haiku tier, Sonnet's input and output rates are twenty times higher; on the larger tier, they are four times higher. That is arithmetic on token rates. It does not tell us the cost per correct fix, how much reasoning each model uses, or how quickly a subscription allowance disappears.
I covered that last distinction in the Sol-versus-Astra usage-limit article. A rate card is one input to a working decision. A result I can accept is the denominator I care about.
A nod to everyone tired of the Codex reset game
There is visible frustration in the current Codex discussions. The October 1 reset thread includes users talking about leaving for Claude. In the October 2 “Reset just hit” thread, some people celebrate, while others say they used a banked reset shortly before the broader reset arrived. The October 7 discussion includes both appreciation for another banked reset and attempts to work out when to spend it.
Those are user reports, not a representative survey or verified account records. They do show why the reset game gets old. People are making decisions about work around announcements, expiry dates and a meter they are trying to interpret.
I can appreciate a reset and still want predictable access. Having to decide whether to pause a useful task in case a reset arrives adds its own cost. A paying developer should be able to understand the allowance in the product and plan a day around it.
OpenAI documents reset banking, and its current usage guidance explains that Work and Codex share usage, with consumption affected by model choice, context, reasoning, tools, retrieval and caching. API prices do not directly specify included subscription tasks. That explains why a simple token-price comparison cannot settle the argument; it leaves room for clearer accounting and communication.
A quick reply when Codex goes down still matters. Predictable capacity matters after the reply, too. These competing releases give frustrated users more options to evaluate, even though no provider's launch establishes that its own limits will suit their workload.
Competition has to reach the working day
Muse Wonder gives me a reason to compare another provider on real tasks. Sonnet offers a stronger case for efficient everyday work. Haiku makes small, carefully bounded jobs cheaper to attempt.
The useful outcome would be more work finished with less supervision, at a cost and cadence I can understand. I would track accepted changes, elapsed time, retries, review effort and allowance consumed on the same representative tasks.
A good coding assistant earns its place by helping me ship. I would rather spend the afternoon checking a working change than checking whether somebody has announced another reset.
What I verified
I read the supplied video's English automatic captions, including the timestamped passages above, and checked the linked Meta, Anthropic and OpenAI pages and Reddit discussions on October 13. Automatic captions can mishear model names; official pages supply the names and prices used here. This article does not independently validate Theo's experiments, vendor benchmark results, individual usage reports or subscription capacity. The illustration is conceptual. Readers should recheck current pricing, access and their account's usage dashboard before choosing a plan.