What Anthropic actually shipped
On June 30, 2026, Anthropic released Claude Sonnet 5, the newest model in its mid-tier Sonnet line and, in the company's words, "the most agentic Sonnet model yet." The headline is not a single benchmark — it's positioning. Sonnet 5 is built to plan, call tools like browsers and terminals, and run multi-step tasks autonomously at a level that a few months ago needed a larger, pricier model. Anthropic also notes it checks its own output without being explicitly asked, a small change that matters a lot for unattended agent loops.
The model ships as the default for Free and Pro plans and is available to Max, Team, and Enterprise users, inside Claude Code and Cowork, and on the API under the model ID claude-sonnet-5. It carries a 1M-token context window — the same as Opus 4.8 and Sonnet 4.6 — so the upgrade is about capability and price, not context length. One launch customer, Zimu Li, summed up the practical pitch: "Claude Sonnet 5 gives our agents a strong execution layer for multi-step software engineering work." For background on the full Claude family, see our Anthropic provider deep-dive.
The model ships as the default for Free and Pro plans and is available to Max, Team, and Enterprise users, inside Claude Code and Cowork, and on the API under the model ID claude-sonnet-5. It carries a 1M-token context window — the same as Opus 4.8 and Sonnet 4.6 — so the upgrade is about capability and price, not context length. One launch customer, Zimu Li, summed up the practical pitch: "Claude Sonnet 5 gives our agents a strong execution layer for multi-step software engineering work." For background on the full Claude family, see our Anthropic provider deep-dive.
Pricing: the August 31 promo, and what it really costs
Sonnet 5 launched with introductory pricing of $2 per million input tokens and $10 per million output tokens, in effect through August 31, 2026. On September 1 it reverts to the standard Sonnet rate of $3 input / $15 output — the same price Sonnet 4.6 charges today. That is the asterisk in the table above: budget for the standard rate if your project outlives the summer, and treat the discount as a window to migrate and load-test cheaply.
Even at the standard rate, Sonnet 5 is dramatically cheaper than the models it rivals on quality. Opus 4.8 runs $5 / $25 per million tokens and Fable 5 — Anthropic's top Mythos-class model — runs $10 / $50. Put numbers on it: a workload of 2M input and 1M output tokens per day costs about $21/day on standard Sonnet 5, $35/day on Opus 4.8, and $70/day on Fable 5. During the promo, Sonnet 5 drops to roughly $14/day. To model your own token mix instead of this toy example, use the API cost estimator or convert a flat budget into tokens with the token-to-USD tool. You can also line all four up side by side on the price comparison tool.
Even at the standard rate, Sonnet 5 is dramatically cheaper than the models it rivals on quality. Opus 4.8 runs $5 / $25 per million tokens and Fable 5 — Anthropic's top Mythos-class model — runs $10 / $50. Put numbers on it: a workload of 2M input and 1M output tokens per day costs about $21/day on standard Sonnet 5, $35/day on Opus 4.8, and $70/day on Fable 5. During the promo, Sonnet 5 drops to roughly $14/day. To model your own token mix instead of this toy example, use the API cost estimator or convert a flat budget into tokens with the token-to-USD tool. You can also line all four up side by side on the price comparison tool.
Sonnet 5 vs Sonnet 4.6: the generational jump
The clearest way to see what changed is to compare Sonnet 5 with the model it replaces. On Anthropic's reported agentic-coding evaluation, Sonnet 5 scores 63.2% versus Sonnet 4.6's 58.1% — a meaningful step up on the kind of multi-file, tool-using coding that agents actually do. The gap is even wider on knowledge work, where Anthropic's Elo-style score jumps from 1395 to 1618. That is not an incremental refresh; it is a different class of model at the same nominal price tier.
Sonnet 5 is also safer in agentic settings. Anthropic reports an overall lower rate of undesirable behaviors than Sonnet 4.6, with better refusal of malicious requests, stronger resistance to prompt-injection attacks, and reduced hallucination and sycophancy. For anyone running Sonnet 4.6 in production, the upgrade path is straightforward — same price after the promo, same 1M context, better numbers across the board. See the head-to-head: Sonnet 5 vs Sonnet 4.6 for the side-by-side.
Sonnet 5 is also safer in agentic settings. Anthropic reports an overall lower rate of undesirable behaviors than Sonnet 4.6, with better refusal of malicious requests, stronger resistance to prompt-injection attacks, and reduced hallucination and sycophancy. For anyone running Sonnet 4.6 in production, the upgrade path is straightforward — same price after the promo, same 1M context, better numbers across the board. See the head-to-head: Sonnet 5 vs Sonnet 4.6 for the side-by-side.
Sonnet 5 vs Opus 4.8: how close is it really?
This is the comparison Anthropic is leaning into, and the data backs the framing. On knowledge work the two are effectively tied — Sonnet 5 scores 1618 to Opus 4.8's 1615 — yet Sonnet 5 costs 40% less per token at the standard rate, and 60% less during the launch promo. For research synthesis, analysis, drafting, and most professional knowledge tasks, that makes Sonnet 5 the obvious default.
Opus 4.8 still wins where it counts most, though. On agentic coding it leads 69.2% to 63.2%, a six-point margin that shows up on the hardest, longest engineering tasks — deep refactors, sprawling codebases, and problems where one extra correct decision per run compounds. Anthropic also notes Sonnet 5 trails Opus 4.8 on some misaligned-behavior metrics, so Opus remains the pick for the highest-stakes autonomous work. The pattern across agentic search (BrowseComp) and computer use (OSWorld-Verified) is the same: Sonnet 5 lands between Sonnet 4.6 and Opus 4.8, much closer to the top than its price suggests. Compare them directly on Sonnet 5 vs Opus 4.8.
Opus 4.8 still wins where it counts most, though. On agentic coding it leads 69.2% to 63.2%, a six-point margin that shows up on the hardest, longest engineering tasks — deep refactors, sprawling codebases, and problems where one extra correct decision per run compounds. Anthropic also notes Sonnet 5 trails Opus 4.8 on some misaligned-behavior metrics, so Opus remains the pick for the highest-stakes autonomous work. The pattern across agentic search (BrowseComp) and computer use (OSWorld-Verified) is the same: Sonnet 5 lands between Sonnet 4.6 and Opus 4.8, much closer to the top than its price suggests. Compare them directly on Sonnet 5 vs Opus 4.8.
Where Fable 5 still earns its premium
Fable 5 sits above the whole Sonnet/Opus stack as Anthropic's Mythos-class flagship, priced at $10 / $50 per million tokens — five times Sonnet 5's standard input cost. Anthropic didn't publish Fable 5 numbers on the Sonnet 5 launch chart, which is itself telling: Fable 5 isn't the model Sonnet 5 is chasing. It's the ceiling, built for the most demanding autonomous knowledge work and the longest-horizon coding agents where capability matters more than the bill.
In practice the lineup now reads as a clean ladder: Sonnet 5 for the bulk of agentic and coding work, Opus 4.8 when you need a few more points on the hardest coding and the strongest safety profile, and Fable 5 for frontier tasks where you'd rather pay 5x than risk a miss. The fact that Sonnet 5 now matches Opus 4.8 on knowledge work is exactly what pushes the premium tiers up-market — they have to justify the gap on the genuinely hard problems.
In practice the lineup now reads as a clean ladder: Sonnet 5 for the bulk of agentic and coding work, Opus 4.8 when you need a few more points on the hardest coding and the strongest safety profile, and Fable 5 for frontier tasks where you'd rather pay 5x than risk a miss. The fact that Sonnet 5 now matches Opus 4.8 on knowledge work is exactly what pushes the premium tiers up-market — they have to justify the gap on the genuinely hard problems.
Which one should you pick?
Default to Sonnet 5 for almost everything: production agents, coding assistants, RAG and long-document work, customer-facing chat, and high-volume pipelines. It is the best quality-per-dollar Claude has shipped, and the August promo makes the switch cheap to test. Move up to Opus 4.8 only for the hardest agentic coding or when its stronger alignment profile is worth the 40% premium, and reach for Fable 5 on frontier reasoning where capability trumps cost.
The most cost-effective setup isn't one model — it's routing. Send the bulk of traffic to Sonnet 5 and escalate only the requests that trip a complexity threshold to Opus 4.8 or Fable 5. Teams that do this typically cut spend 30–50% versus running a flagship for everything, with no quality loss on the paths that matter. Before you commit, price your real token distribution across all four on the API cost estimator, and bookmark our Anthropic pricing guides for how the output multiplier drives the bill.
The most cost-effective setup isn't one model — it's routing. Send the bulk of traffic to Sonnet 5 and escalate only the requests that trip a complexity threshold to Opus 4.8 or Fable 5. Teams that do this typically cut spend 30–50% versus running a flagship for everything, with no quality loss on the paths that matter. Before you commit, price your real token distribution across all four on the API cost estimator, and bookmark our Anthropic pricing guides for how the output multiplier drives the bill.
The competitive backdrop
Sonnet 5 didn't launch in a vacuum. It arrived the same week as a wave of competing agentic releases, including OpenAI's GPT-5.6 "Sol" and Google's Gemini 3.5 Flash — a sign that agentic capability has become table-stakes across every price tier, not a flagship-only luxury. Anthropic's move is to push that capability down into the mid tier and undercut on price, betting that most real agent workloads don't need a flagship.
That's the same bet TokenRate is built around: the right model is rarely the most expensive one, it's the cheapest one that clears your quality bar. Sonnet 5 just moved that bar a lot higher at the mid tier. Compare its live pricing against GPT and Gemini on the price comparison tool, and use the calculator to see what the switch saves on your actual volume.
That's the same bet TokenRate is built around: the right model is rarely the most expensive one, it's the cheapest one that clears your quality bar. Sonnet 5 just moved that bar a lot higher at the mid tier. Compare its live pricing against GPT and Gemini on the price comparison tool, and use the calculator to see what the switch saves on your actual volume.