What Anthropic actually said
The May 28, 2026 announcement that paired Opus 4.8 with the $965B-valuation and $47B-run-rate disclosure was reported, correctly, as the headline business story of the week. The line in the same release that matters more for production AI work was reported as a footnote: Anthropic stated that it expects to bring Mythos-class models to all customers "in the coming weeks."
That sentence has a specific operating meaning. Mythos has been live since April 7, 2026 under Project Glasswing — an invitation-only access program covering twelve founding partner organizations plus roughly forty additional vetted critical-infrastructure operators, with partner pricing of $25/M input and $125/M output tokens (five times the standard Opus 4.7 / 4.8 price card). The model scores 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, 82.0% on Terminal-Bench 2.0, and 97.6% on USAMO 2026 — a score that's 55.3 percentage points higher than Opus 4.6 on the same olympiad. Glasswing was a deliberately small-surface deployment, with the access vetting reflecting Anthropic's caution about a model that, in the company's own published evaluations, demonstrated meaningful capability uplift on cybersecurity tasks including zero-day identification and exploitation when directed to do so.
Coming weeks maps to general availability on the standard API, with the safeguards Anthropic has spent two months building inside Glasswing now deployed on the public surface. The price card may move; the access policy may include an enterprise-only band for a while; specific capability classes (the most concerning agentic deployments) may stay gated longer. But the existence of a Mythos tier on the same API your team already calls — sitting one full capability tier above Opus 4.8 — is the structural change.
Why "one tier above" is a bigger deal than the same-tier upgrades were
For the last eighteen months, every model upgrade your team evaluated against — Opus 4.5 to 4.6 to 4.7 to 4.8, GPT-5 to 5.5, Gemini 3.0 to 3.1 to 3.5 — moved a few percentage points on most benchmarks, with occasional larger jumps on specific capability axes. The eval harness you ran against the new model was structurally similar to the one you ran against the previous one. The routing policy needed minor recalibration. The review workflow stayed roughly the same shape.
Mythos is not that kind of upgrade. The published benchmark spread between Opus 4.6 and Mythos on USAMO is larger than the entire spread between GPT-3.5 and GPT-4 on the same kind of reasoning task. The SWE-bench Verified jump from Opus 4.7's 87% to Mythos's 93.9% is not a better Opus; it's a model that completes a meaningfully different distribution of software-engineering tasks correctly on the first try. The reasoning-task scores are at a level where the failure modes change qualitatively — the model is wrong less often, and when it is wrong, it is wrong in different ways than the model below it was. That is the harder shape of model upgrade to absorb operationally, because every assumption baked into the harness about how this model tends to fail needs to be re-derived.
Three axes where the change is most visible.
The hard-task tail moves into reach. Workloads that the Opus tier completes 60-70% of the time, on first try, with substantial review burden — and that the team has therefore designed around with retrieval scaffolding, manual decomposition, and prompt engineering — move to 85-90% with the new model. The scaffolding that was load-bearing under Opus is now overhead. The decomposition that was a precondition is now sometimes counterproductive.
The agentic-trajectory quality changes shape. Long-horizon agentic runs that previously failed by losing the thread halfway through — the classic the agent forgets the constraint it was given thirty steps ago failure mode — are visibly less common with Mythos. The runs that succeed at the Opus tier mostly succeed faster at Mythos; the runs that failed at Opus sometimes succeed at Mythos, and the cases where they don't are different cases. The eval harness graded against Opus-era failure distributions will systematically misreport which workloads benefit from the upgrade.
The dangerous-capability surface expands. Anthropic's own evaluations document Mythos's improved capability on offensive cybersecurity tasks. Under Glasswing, the deployment surface was small enough and the access gate strong enough that this was a controllable risk. Under general availability, the access gate becomes the routing policy, the prompt-injection defenses, and the review queue inside your application. You now own a piece of the safety surface that, under the Glasswing era, Anthropic owned exclusively.
What needs to be true in your stack before the rollout
Four things to set up this month. None are exotic; all are deferable until the model arrives, at which point you start the work two weeks late on a four-week sprint.
Your eval harness needs a real top row. If the current top of your eval matrix is Opus 4.8, you have one position to compare against until Mythos arrives. The eval discipline that produces honest comparisons needs at least two reference rows above the model under evaluation — a frontier-grade reference that establishes the ceiling, and a tier-down reference that establishes the cost-quality tradeoff. Adding a top row for Mythos is a four-day engineering job done in advance; it's a two-week scramble done after the model lands and product is asking should we switch?
Your routing policy needs a tier above the most expensive current option. Most production routers today treat Opus 4.8 (or GPT-5.5, or Gemini 3.5 Pro) as the most expensive option in the matrix and route to it for the hardest tasks. A Mythos rollout adds a tier that's more expensive again, with a capability bump that justifies it for a specific narrow band of workloads. The router needs an explicit policy for what makes a task Mythos-worthy — the eval criterion for routing escalation — and the budget guardrail that prevents the policy from being abused. Without the policy, the model gets used by whoever happens to remember it exists, which is the worst possible cost distribution.
Your review workflow needs a senior-review queue specifically calibrated for Mythos output. A model that's right 93.9% of the time on SWE-bench Verified is a model whose errors are harder to catch. The 6.1% wrong outputs are statistically more likely to be wrong in subtle, plausible-looking ways than the 13% wrong outputs from Opus 4.8 — because the easy-to-catch errors are exactly the ones the model upgrade eliminated. The senior-reviewer time per Mythos output is going to be higher per flagged output than it was for the equivalent Opus output, because the cases where the model is wrong are systematically the harder-to-detect ones. Staffing the senior-review queue for Mythos quality of error, not Mythos rate of error, is the discipline that matters.
Your prompt-injection and tool-call surface needs a Mythos-specific review. The dangerous-capability uplift documented in Anthropic's Mythos evaluations is most relevant in agentic deployments where the model is taking actions on behalf of a user, with access to external tools, against an attacker-controllable input distribution. The standard adversarial review of what an attacker can do if they own one prompt in the chain gets graded against the most capable model in the routing matrix, not the median one. Before Mythos is wired into any production agent path, the adversarial review needs a refresh — and the most expensive failure modes are the ones the previous tier couldn't reach, so the historical review doesn't cover them.
What the rollout does not change
Three caveats that the next round of marketing on Mythos will work hard to make you forget.
It does not eliminate the multi-vendor portability question. Mythos is a Claude model. GPT-5.5 Pro, Gemini Mythos-equivalents (which the other labs will ship in response within the quarter), and the open-weight models will continue to constitute a real portfolio of options for production workloads. The team that becomes Anthropic-only because Mythos exists is the same team that became Anthropic-only because Opus 4.7 was best-in-class for six weeks last fall — and they will pay the same portability tax when the relative-capability ranking flips again. The eval-and-route discipline that worked across vendors needs to keep working across vendors, with Mythos as a fourth or fifth column in the matrix, not a replacement for it.
It does not eliminate cost discipline. A model priced at $25/M input and $125/M output, even with a capability uplift, is the most expensive token on your routing surface by a wide margin. The cost-per-successful-task math is going to favor Mythos for a small subset of workloads where the capability difference matters; it is going to be a structural loss versus Opus 4.8 for the majority of routine work. The teams that route everything to Mythos because it's the best model are the teams whose AI line item triples without a corresponding triple in productivity.
It does not collapse the human-in-the-loop requirement. A 93.9% correctness rate is a 6.1% incorrect-and-shipped rate if nothing reviews the output. For low-stakes workloads — exploratory analysis, draft generation, scaffolding code — the autonomous-action calculus stands. For anything that ships to a customer, modifies production data, or constitutes legally significant action, the senior-reviewer queue stays in the loop. The model getting better changes the throughput the queue can absorb; it does not change the requirement that the queue exists.
Where Sonnet Code fits
A frontier-tier model rolling out to general availability is the easy half of the story. The hard half is the engineering above the model that turns a new top row on the API into a portable, evaluatable, governable production capability. AI development at Sonnet Code is that engineering: extending your eval harness with a Mythos-calibrated reference row that produces honest cost-quality comparisons across your actual workload, designing the routing policy that escalates the right tasks to the right tier with budget guardrails that prevent the most expensive option from being the default by accident, and refreshing the adversarial review on the prompt-injection and tool-call surface for the capability uplift Mythos brings. AI training is the human-judgment half: senior engineers and domain experts who design the rubrics that distinguish Mythos-worthy tasks from the broader workload, calibrate the senior-review queue for the harder-to-catch failure modes a more capable model produces, and stand up the gold sets that grade Mythos honestly against the tier below it on the workloads that actually matter to your product.
The Glasswing era ends in the coming weeks. The general-availability era begins. The teams that walk into it with the eval, routing, review, and adversarial-review layers already calibrated will compound on the new top row from week one. The teams that walk into it cold will spend the next quarter discovering, in production, which of their assumptions about how this model fails needed to be re-derived from scratch.
