Open “Article briefing” to copy it manually, then choose an assistant and paste it there. Use an assistant you are allowed to use for work, and leave out personal or confidential information.
FSR BRIEFING PACKET - Claude Sonnet 5.5: What Changed, Who Can Use It, and When to Consider Opus 5.5
Publisher: Future Stack Reviews. Source: https://future-stack-reviews.com/claude-sonnet-5-5-starter-tierc/
Format: Starter Guide. Evidence depth: Tier C, built from Anthropic's public pages and one independent evaluator's published results; FSR ran no tests of Sonnet 5.5. Read 29 to 30 SEP 2026, Asia/Tokyo. No affiliate or paid links. FSR used AI assistants, including Anthropic's Claude, to research and draft the article.
WHAT IT IS
Sonnet 5.5, released September 28, 2026, is Anthropic's newest Sonnet model. Anthropic positions it for well-scoped everyday work and says Opus 5.5 remains clearly stronger at complex, open-ended work that needs sustained judgment. Anthropic's Sonnet page says anyone can chat with Sonnet 5.5 on Claude.ai. The Free plan does not include Opus.
KEY NUMBERS
- API list price per million tokens: Sonnet 5.5 and Sonnet 5, $2 input and $10 output; Opus 5.5, $4 and $20.
- Anthropic: up to 30% less per task than Sonnet 5 on most work, in its testing, and output 30%+ faster.
- Artificial Analysis, an independent evaluator (API benchmark runs; release pages read September 30, 2026; cost per index task / index score). Medium effort: Sonnet 5.5 $0.59 / 41, Sonnet 5 $1.00 / 28, Opus 5.5 $1.34 / 51. Max effort, same order: $7.60 / 56, $5.09 / 38, $5.98 / 58. It concludes that, for the cost, Sonnet 5.5 sits behind Opus 5.5 at its highest effort levels. It tested a pre-release build with a bug that Anthropic says is now fixed; a re-run is planned. Not app charges.
- In the apps, Anthropic lists Medium as Sonnet 5.5's default effort; check the level shown. Higher effort reaches usage limits faster. No Anthropic page FSR read says how much of a plan's limit each model uses.
WHAT TO CHECK
- The model and effort next to the send button. No Anthropic page FSR read says which model a new chat starts on, by plan. On Enterprise, an administrator can turn models or effort levels off.
- Some flagged cybersecurity and AI-development requests switch to Sonnet 5, with a notice and a model label on the reply. The checks can misfire, and Sonnet 5 can also block a request.
- Data: Anthropic's terms and pricing page call training on Free, Pro and Max chats opt-out; its privacy article says you choose to allow it. Check "Help Improve our AI models" in Settings > Privacy before you paste anything that matters. Team and Enterprise content is not used for training by default. Incognito chats are not used for training but are kept 30 days by default.
- Pro and Max: if you enable and fund usage credits, you can choose to keep working past your limit at standard API rates. Chats and Claude Code share the same limits.
A SMALL FIRST TEST (FSR proposal, not a test FSR ran)
Number the paragraphs of a public or non-confidential text of 250 to 400 words, and choose three facts, one limitation and an audience. Note how your current method handles it. In a new chat, check that Sonnet 5.5 is selected, use Medium if offered and note the effort shown. Ask for a customer update of no more than 120 words that uses only the text, includes the four points, adds nothing unsupported and lists the paragraph behind each sentence. Check every sentence yourself; the list is a pointer, not proof. Log the time, retries and the model named on each reply. Do not count a reply labeled Sonnet 5, and stop at a usage-limit notice. One run is not a verdict on the model; if your current method did better, keep it.
FSR VERDICT
If Sonnet 5.5 is in your model menu, try this test once, at the default effort. For open-ended work that needs sustained judgment, consider Opus 5.5 first if your plan includes it. Keep a workflow that already meets your standard, and keep a person reviewing consequential work you cannot check. Per-model usage figures from Anthropic, or a re-run on the fixed build, could change this verdict.
SOURCES
- Launch post: https://www.anthropic.com/claude-sonnet-5-5
- Model and effort settings: https://support.claude.com/en/articles/8664678-change-the-model-effort-and-thinking-settings
- Model switching: https://support.claude.com/en/articles/17161993-why-claude-switched-models-in-your-conversation-with-sonnet-5-5
- Privacy setting: https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings
- Usage credits: https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plans
- API pricing: https://platform.claude.com/docs/en/about-claude/pricing
- Artificial Analysis: https://artificialanalysis.ai/models/releases/claude-sonnet-5-5
LIMITS
Prices, plans and policies change; where this packet and the live pages disagree, the live pages win. Nothing here measures output quality, speed or cost, or shows what a particular account can access.
Help me decide whether to try Sonnet 5.5, keep my current setup or use Opus 5.5 for a task. Ask which plan I am on, whether an employer manages my account, what work I want to do and how I would check the result. Do not ask me to paste confidential, client or personal material.
Your assistant's reply is its own answer, not FSR's. Check it against the sources above.
By Takashi Fujino
Evidence read September 29 to 30, 2026
TierC
Evidence record: Starter Guide · Tier C — Document-based Research
Built from Anthropic's public pages and one independent evaluator's published results. FSR ran no tests of Sonnet 5.5.
Starter Guide is the format; Tier is depth of evidence, not a rating.
Read
SEP 30 2026Last read September 30, 2026Asia/Tokyo
Plan figures
API prices are Anthropic's US dollar list prices. No app plan prices are quoted.
Paid links
None on this page.
Claude Sonnet 5.5 is the newest model in Anthropic's Sonnet line, released September 28, 2026. Anthropic aims it at well-scoped everyday work, says it generates output faster than Sonnet 5 at the same API price, and keeps Opus 5.5 for complex work that needs sustained judgment.
If Sonnet 5.5 is in your model menu, it's worth one small test at the default effort, on writing you can check against a source. For open-ended work that needs sustained judgment, consider Opus 5.5 first if your plan includes it. Highlighted: The speed and savings claims are Anthropic's. In an independent API benchmark of a pre-release build, Sonnet 5.5 cost less per task than Sonnet 5 and Opus 5.5 at Medium effort but more at Max, and scored between Sonnet 5 and Opus 5.5 at both. FSR has not tested it.
Sources: Anthropic launch post · Artificial Analysis release pages for Sonnet 5.5, Sonnet 5 and Opus 5.5, cost per Intelligence Index task and index score by effort (read September 30, 2026)
Best for, not for, and when to wait
Best for · Not for · Wait
Best for
If all apply
Claude app users who write or analyze, alone or at work
01who see Sonnet 5.5 in their model menu
02who have a writing task they can check against a source, such as a customer update
03Highlighted: who will review what it writes before anyone relies on it
Not for
If either applies
01Consequential work that no qualified person will check
02Material you may not share, whether it's yours, an employer's or a client's
Wait
If either applies
01An employer manages your Claude account, and its rules on data or models are unclear
02You run Sonnet 5 through the API and haven't read the migration changes
Anthropic claims 30%+ faster output than Sonnet 5; FSR hasn't measured it.
Who can use it
Anyone on Claude.ai, per Anthropic's Sonnet page.
Free doesn't include Opus. The starting model for new chats isn't stated by plan.
API price
$2 input and $10 output per million tokens, the same as Sonnet 5.
Opus 5.5: $4 and $20.
"Up to 30% less"
Per task, against Sonnet 5, in Anthropic's own tests.
Red note: Independent API benchmark, pre-release build: $0.59 per task at Medium effort, against $1.00 for Sonnet 5 and $1.34 for Opus 5.5; at Max, $7.60 against $5.09 and $5.98. Not app charges.
In the apps
Medium effort by default, per Anthropic; check the level shown. Thinking stays on.
Higher effort reaches your usage limits faster.
Model switching
Some flagged cybersecurity and AI-development requests go to Sonnet 5 instead.
A notice and a model label on the reply show it.
Your data
On Free, Pro and Max, check the training setting before sharing; Anthropic's pages describe it differently. Team and Enterprise content isn't used for training, by default.
Red note: Incognito chats are still kept 30 days by default.
Anthropic pitches Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5, strongest at well-scoped everyday tasks, bug fixes, and creating documents, slides and spreadsheets. It says Sonnet 5.5 generates output 30%+ faster than Sonnet 5; the launch post gives no whole-task times, and FSR hasn't measured either.
Unchanged from Sonnet 5, per Anthropic's API docs: the API price, the tokenizer, the 1 million-token context window and the 128,000-token output limit, which describe the API rather than any app plan. New: a reliable knowledge cutoff of June 2026, against January 2026 for Sonnet 5, and cybersecurity fallbacks, a first for a Sonnet model (section 06).
Anthropic's Sonnet page says anyone can chat with Sonnet 5.5 on Claude.ai. Anthropic's plan table and app tutorial put the Sonnet line on the Free plan and Opus on paid plans, without saying whether that includes Sonnet 5.5.
No Anthropic page FSR read says which model a new chat starts on, by plan, so check before you type. The model name and effort level sit next to the send button; click the name to switch, and look under More models if Sonnet 5.5 isn't listed. On Enterprise, an administrator can turn models or effort levels off and set the model new chats start on.
Anthropic lists Medium as the default effort for Sonnet 5.5 in the apps, with thinking on, but check the level shown. Higher effort is slower and reaches your usage limits faster.
Sonnet 5.5 costs the same per token as Sonnet 5, so Anthropic's saving is per task: up to 30% less than Sonnet 5 on most work in its testing, because it tends to need fewer tokens.
Artificial Analysis, an independent evaluator, runs its own benchmarks through the API. On September 30, its release pages showed Sonnet 5.5 at Medium costing $0.59 per task, against $1.00 for Sonnet 5 and $1.34 for Opus 5.5, and scoring 41 on its index, against 28 and 51. At Max the cost order flipped: $7.60, against $5.09 and $5.98, with scores of 56, 38 and 58. Its Medium figures point the same way as Anthropic's claim; its Max figures, the other way. It tested a pre-release build that had a bug. Anthropic says the bug is fixed and expects it, if anything, to have lowered Sonnet 5.5's scores slightly. Artificial Analysis plans a re-run; its pages on September 30 didn't say whether it had happened.
None of these figures is your bill. No Anthropic page FSR read says how much of a plan's limit each model uses; Anthropic's app tutorial rates the Sonnet family as moderate on limits and Opus as heavy. On Pro and Max, if you enable and fund usage credits, you can choose to keep working past your limit at standard API rates; chats and Claude Code share the same limits.
Anthropic says Sonnet 5.5 at Max performs comparably to Opus 5.5 on several evaluations, but that Opus 5.5 remains clearly stronger at complex, open-ended work that needs sustained judgment. Its app tutorial says to start with Sonnet when unsure. Artificial Analysis concludes that, for the cost, Sonnet 5.5 sits behind Opus 5.5 at its highest effort levels.
Anthropic's consumer terms and pricing page call training on Free, Pro and Max chats opt-out, while its privacy article says you choose to allow it, and none of them says how a new account starts. Check "Help Improve our AI models" in Settings > Privacy before you paste anything that matters. Turning it off keeps your new and previously stored chats out of future training, but not out of training already started; chats flagged for safety review, and feedback you send, follow separate rules. Team and Enterprise content isn't used for training by default, though feedback can be; Anthropic keeps feedback for up to five years.
Incognito chats aren't used for training or saved to your history or memory. Anthropic still keeps them for 30 days by default, longer under an Enterprise retention setting, and on Team and Enterprise they're included in the organization's data exports. Follow any employer or client rules first.
Anthropic says most requests never trigger its safety checks, but some flagged ones switch to Sonnet 5 in the same chat: higher-risk offensive security work, such as writing an exploit, may fall back, and flagged work on building the most advanced AI models does. Secure coding, such as scanning source code for vulnerabilities, stays on Sonnet 5.5, though the checks can misfire. Sonnet 5 can also block a request, and requests that could aid serious biological harm or extract the model's internal reasoning are blocked outright.
The checks cover everything the model reads, including memory, connector content, web results and files. A notice appears, the reply is labeled with the model that answered, and the menu stays on Sonnet 5 until you switch back. While the flagged request remains, switching back can bring another fallback, so Anthropic suggests editing that message. Turning off "Switch models when a message is flagged" in Settings > Capabilities pauses the chat instead.
FSR hasn't run this exercise. Use only text you're allowed to share (see section 05), and on Pro or Max, check first whether usage credits are enabled.
Don't count a reply labeled Sonnet 5 as a Sonnet 5.5 result, and stop at a usage-limit notice; on Pro or Max, going on can draw on paid credits. One failed run is a result for this task, not a verdict on the model, and if your current method did better, keep it.
Claude Code still defaults to Opus 5.5 on Pro, Max, Team, Enterprise and the API, and Anthropic's developer guides start most API workloads on Opus 5.5. Sonnet 5.5 needs Claude Code 2.1.284 or later: run claude update, then select it with /model and its full name, claude-sonnet-5-5. If an ANTHROPIC_API_KEY environment variable is set, Claude Code bills the API, not your plan.
On the API, moving from Sonnet 5 to claude-sonnet-5-5 takes more than a name change: Anthropic lists five breaking changes, and a request that turns thinking off now returns an error; switch it to the new between_tools setting, which works only at low, medium or high effort. API fallback to Sonnet 5 is opt-in; until you configure it, a flagged request returns a refusal stop reason instead of an answer. Sonnet 5 stays available as a legacy model until at least June 30, 2027.
Anthropic's Sonnet page says anyone can chat with it on Claude.ai, and the Free plan includes the Sonnet line but not Opus. Which model a new Free chat starts on isn't stated, so check the model menu.
Not by Anthropic's account: it rates Opus 5.5 clearly stronger at complex, open-ended work, and Opus led on seven of eight launch benchmarks, most by small margins. On the API, Sonnet 5.5 costs half as much per input and output token.
No Anthropic page FSR read gives a figure by model. "Up to 30% less" is Anthropic's cost-per-task comparison with Sonnet 5, not a share of your plan's limit; its app tutorial rates the Sonnet family as moderate on limits and Opus as heavy.
How often does Claude switch models, and how can I tell?
No Anthropic page FSR read gives an everyday rate; in Artificial Analysis's API benchmark runs, about 0.1% of tasks fell back. A switch shows a notice and labels the reply with the model that answered.
Starter Guide, Tier C. FSR ran no tests of Sonnet 5.5 or Opus 5.5 for this briefing, so nothing here measures their quality, speed or cost on real work. Section 07 is a proposal.
FSR read the pages cited here on September 29 and 30, 2026 (Asia/Tokyo), and on September 30 an FSR editor also viewed the pricing page, the top of the Sonnet page and Artificial Analysis's Sonnet 5.5 and Opus 5.5 model pages. From Japan, the pricing page showed Japanese, tax-inclusive prices, so no plan prices are quoted. The Sonnet page still carries older copy; FSR used only its Sonnet 5.5 availability line. Statements that a page is silent cover only the pages listed.
FSR researched and drafted this briefing with AI assistants, including Anthropic's Claude, the model family it covers, and checked their findings against the pages cited; agreement between AI tools is not independent verification.
Highlighted: If Sonnet 5.5 is in your model menu, try the section 07 test once, at the default effort. For open-ended work that needs sustained judgment, consider Opus 5.5 first if your plan includes it. Keep a workflow that already meets your standard, and keep a person reviewing consequential work you can't check.
The case for a test rests on Anthropic's claims of faster output and lower cost per task than Sonnet 5. Artificial Analysis's per-task costs agree at Medium effort but not at Max, and its index places Sonnet 5.5 between Sonnet 5 and Opus 5.5 at both. Per-model usage figures from Anthropic, or a re-run on the fixed build, could change this verdict.
For readers whose employer runs Claude: how the plans differ on contracts, training, audit logs, storage and spending. Read in August, before the 5.5 models.
What decides the model and the bill in a Claude Code session, including an API key left in your shell. Read in August; for current model defaults, see this briefing's section 08.
Starter Guide. Evidence read September 29 to 30, 2026 (Asia/Tokyo). Tier C: FSR ran no tests of Sonnet 5.5. No affiliate or paid links. Prices, plans and policies change; check Anthropic's pages before you rely on them.
Read up to five eligible articles in full per calendar month in this browser. Enable a random, essential reading cookie to use this allowance. Clearing cookies or changing browsers resets this browser allowance.