Models, context, and what your seat actually buys

TL;DR. For 90% of Cowork work, Sonnet 4.6 is the right model — pin it as the default and stop thinking about it. Override to Opus only for legal, regulatory, or deep multi-document analysis. Haiku is for bulk simple tasks. The 1M-token context window matters less than people think; memory and a CLAUDE.md matter more. Procurement-wise, plan on a mix of Pro and Max seats per team — not one tier across the board.

The default-first principle#

Cowork users should not be choosing models in the moment. They should be doing the work.

Pin Sonnet 4.6 as the workspace default. It runs Cowork's agentic loop at near-Opus quality for roughly one-fifth the quota cost. It handles document work, multi-step reasoning, OCR, and almost every workflow in this bible.

Override to a different model only when the work is one of the named exceptions in the next section. Most teams discover after a quarter that they have never overridden the default — which is the right outcome.

The three models, in plain English#

The relative-cost numbers are honest as of April 2026 and shift when Anthropic publishes new ratios. Treat them as orders of magnitude, not exact prices.

The Opus exceptions#

Three patterns where Opus earns its quota multiplier:

• Legal or regulatory analysis. Contract diff with reasoning about clause-level risk; regulatory mapping (e.g., GDPR vs CPRA) where the work is explanation, not lookup.
• Long-form scientific or technical synthesis. Research literature reviews, RFC analysis, white-paper synthesis — anything where the value is in the connections the model makes across hundreds of pages.
• Multi-agent coordination. When one Cowork run kicks off sub-agents that need to coordinate complex outputs, Opus's planning quality is worth the cost.

Outside these three patterns, Sonnet wins on speed and quota at quality almost nobody can tell apart in practice. We have run blind comparisons with finance and ops teams; the consensus is that Opus is "noticeably more careful" on the hardest 5% of tasks and "indistinguishable" on the rest.

The 1M context window — what it actually means#

Sonnet 4.6 and Opus 4.6 support a one-million-token context window (in beta) and 128K output. In English: roughly 1,500 pages of material can sit inside a single Cowork session.

The practical implication for mid-market is real but narrower than the marketing suggests. A quarter of supplier PDFs, a 200-page contract, or a year of monthly reports can be analysed in one run. That genuinely changes the workflow for legal and finance teams whose work used to require chunking by hand.

The caveat: long context costs more quota and is sometimes overkill. We recommend chunking when the work is naturally chunkable — summarise each invoice, then aggregate is faster, cheaper, and more accurate than here are 200 invoices, give me the summary. Use the big window when the work genuinely needs cross-document reasoning, not by default.

Memory: the other "context"#

Cowork now persists memory across sessions — a capability that rolled out late 2025 and matured through early 2026. Memory survives restarts; it does not survive account changes.

The Tinkso convention is to keep a CLAUDE.md file in the workspace folder. It acts as the team's contract with Cowork: house style, file naming, recurring formats, things to never do. Cowork reads it on every session and treats its contents as standing instructions. This is more reliable than memory alone — a written file can be reviewed, version-controlled, and shared across the team. Memory is per-user; CLAUDE.md is per-workspace.

Full pattern is on CLAUDE.md and memory.

The quota math, for procurement#

Three plans, three usage envelopes. Verify against current Anthropic pricing before you publish a budget.

• Pro ($20/month). Roughly one to one-and-a-half hours of intensive Cowork work per five-hour reset window on Sonnet. Resets every five hours.
• Max ($100–200/month). Approximately 5× the Pro envelope. Right plan for daily power users — finance analysts, marketing operators, people whose week revolves around Cowork.
• Team / Enterprise. Org-level quotas, SSO, admin console, audit logs. Required for any deployment over ten seats and for any team that needs SCIM or compliance controls.

The advice we give consistently: pilot teams on Pro for two weeks, measure actual consumption, then pick the seat mix. Buying everyone Max from day one over-spends; buying everyone Pro under-equips your power users.

Mixed seats are normal#

Most mid-market deployments end up with three seat tiers, in roughly these proportions:

• Pro — occasional users (~50% of seats)
• Max — daily heavy users (~30% of seats)
• Team / Enterprise admin seats — IT, governance, sponsors (~20% of seats)

This is the right shape, not a sign you over- or under-bought. The shape shifts toward more Max seats over time as power users emerge. Plan to revisit the mix at the end of every quarter.

Tinkso's take#

We watch first, license second. The shadow week tells us who is a heavy user and who is not, before HR or Procurement has to lock in seat counts. Three weeks of usage data is worth more than three weeks of debate about plan tiers.

We rarely recommend Opus as a default. The cost-quality math does not justify it for the work mid-market teams put through Cowork. When a client insists on Opus across the board, we ask them to run one week of A/B comparisons on a sample of their own work — they almost always come back saying "keep us on Sonnet".

Try this#

In the Cowork settings, set Sonnet 4.6 as your default. Run one full week without changing it. Note any moment where you feel quality is missing — those are your candidates for an Opus override. Most weeks you will find none.