Claude Sonnet 5 is the most agentic Sonnet model yet. It closes the gap with Opus 4.8 on reasoning, tool use, coding, and knowledge work at roughly half the cost. For GTM engineers running agents in Claude Code, Hermes, or OpenClaw, it's the default model for everyday work.
Source: anthropic.com/news/claude-sonnet-5 (June 30, 2026). Model string: claude-sonnet-5. API docs.
The single most useful idea here is the routing strategy in the second section: plan with Opus, execute with Sonnet 5. That alone cuts roughly half the cost off most agent workloads.
What Sonnet 5 is and what changed
Anthropic's own product team is using this model for 65% of their code.
What changed from Sonnet 4.6
Sonnet 5 is a full number upgrade, not a point release.
- Reasoning — substantially better multi-step reasoning. It checks its own output without being asked, and doesn't stop short on complex tasks the way Sonnet 4.6 did.
- Tool use — better at sustained tool use across messy technical contexts. Follows through on multi-step workflows where previous Sonnet models would stall halfway.
- Coding — handles brownfield code well: race conditions, hidden tests, legacy code. Traces failures to root cause rather than patching symptoms. In testing it wrote a reproducing test, implemented the fix, then stashed the fix to confirm the bug returned without it — all in a single pass, without being asked.
- Knowledge work — slightly better than Opus 4.8 on knowledge work benchmarks.
- Speed — significantly faster than Opus 4.8.
- Safety — lower rate of undesirable behaviours than Sonnet 4.6, lower hallucination rate, lower sycophancy rate, and better at refusing malicious requests and resisting prompt injection in agentic contexts.
Sonnet 5 vs the field
| Model | Input cost | Output cost | Agentic capability |
|---|---|---|---|
| Sonnet 5 (intro pricing, through Aug 31 2026) | $2/MTok | $10/MTok | Close to Opus 4.8 |
| Sonnet 5 (standard pricing, from Sep 1 2026) | $3/MTok | $15/MTok | Close to Opus 4.8 |
| Opus 4.8 | $5/MTok | $25/MTok | Highest capability |
| Sonnet 4.6 | Standard | Standard | Falls well short of Opus 4.8 |
At introductory pricing, Sonnet 5 is roughly cost-neutral versus Sonnet 4.6 — the updated tokenizer processes text differently, mapping the same input to roughly 1.0–1.35x more tokens depending on content type.
At standard pricing, Sonnet 5 costs about half what Opus 4.8 costs for nearly comparable agentic performance. For GTM engineers spending thousands per month on Claude tokens in Hermes or OpenClaw, that brings the bill down significantly while keeping performance close to Opus.
Where to use it, and where not to
Use Sonnet 5 for:
- Everyday coding tasks in Claude Code
- All agent execution work in Hermes and OpenClaw
- Multi-step GTM workflows: account research, sequence execution, CRM updates, inbox triage
- Anything where Sonnet 4.6 was good enough but Opus felt like overkill
- Anything where cost matters and you want near-Opus quality
Use Opus 4.8 for:
- Complex planning sessions in Claude Code (plan mode, ultra code mode)
- Spinning up multi-agent workflows with sub-agents
- Hard reasoning tasks where you need maximum compute
- Cybersecurity work that requires reduced guardrails (Opus has the Cyber Verification Program)
Don't use either for:
- Developing software exploits — neither Sonnet model can do this by design
- Tasks where Claude Mythos Preview is needed (restricted access, not public)
Pricing, setup, and model routing strategy
Sonnet 5 is the new default for GTM agents running through Claude Code, Hermes, and OpenClaw. If you were using Opus for cost reasons, switch. If you were using Sonnet 4.6, upgrade immediately.
Switch your model string
API model string: claude-sonnet-5
In Claude Code Desktop, use the bottom-right dropdown and select Sonnet 5.
In Hermes or OpenClaw via API, if Sonnet 5 isn't yet in your model directory:
Switch the Claude API model string to claude-sonnet-5. Look up the current API string format online and update it in the backend configuration.
In Claude Code via plugin or settings:
// In your .claude/settings.json or equivalent config
{
"model": "claude-sonnet-5",
"effort": "medium"
}
The recommended model routing strategy
Use Opus 4.8 for planning, Sonnet 5 for execution.
Why it works: if the plan is detailed and accurate, you don't need the most powerful model in the world to execute it. Sonnet 5 executing a well-structured Opus plan produces the same quality output for significantly less cost.
Step 1 — planning. Open plan mode in Claude Code Desktop. Switch to Opus 4.8, and to Ultra Code mode (or Max if you're not on the 20x plan).
Go into plan mode, ultra code. I want to [BUILD THIS THING / PLAN THIS WORKFLOW].
Opus spins up sub-agents (Workflows) to design the architecture — multiple agents, lots of tokens, lots of tool use. This is where you spend the compute budget.
Step 2 — execution. After plan mode completes and writes a full markdown plan, switch to Sonnet 5 at medium effort.
Execute on the plan.
Because the plan is already structured, medium effort is sufficient. Result: near-Opus quality output at Sonnet-level cost for the 80% of tokens that go into execution.
Effort levels and when to use each
| Effort level | Use for | Cost |
|---|---|---|
| Low | Simple tasks, quick queries, basic formatting | Lowest |
| Medium | Most GTM agent work: research, writing, CRM updates | Moderate |
| High | Complex reasoning, hard debugging, multi-step agentic workflows | Higher |
| Extra high (xhigh) | Maximum capability, approaching Opus 4.8 on some tasks | Highest |
For most GTM workflows, medium is the right setting. Use high or xhigh when the task is genuinely hard and medium is producing incomplete results.
Agentic GTM workflows: what Sonnet 5 handles best
Sonnet 5 is primarily a model for agentic work — multi-step workflows, sustained tool use, autonomous execution.
What the early access partner data says
Direct quotes from Anthropic's early access partners on what Sonnet 5 does better than Sonnet 4.6:
Handles sustained coding, tool use, and debugging well across messy technical contexts. Especially useful for workflows where follow-through and technical grounding matter.
We handed it a two-part job: update Salesforce account tiers, send a launch announcement to enterprise contacts. It finished end to end. That used to stall halfway. For day-to-day automation, it's a no-brainer.
Gets more done with less. Same output quality, fewer steps to get there.
Agents stay on plan, follow our conventions, and ship clean multi-step changes.
Traces a failure to its actual root cause and ships a durable fix instead of patching the symptom.
The pattern across all of these: it finishes things. Where Sonnet 4.6 would stop short, Sonnet 5 follows through.
Multi-step CRM automation
Run a CRM hygiene pass.
1. Pull my external meetings from the last 2 weeks
2. Match against existing CRM accounts using fuzzy domain logic
3. Flag contacts that attended meetings but are not in CRM
4. Flag accounts with no open opportunity
5. Show me the report before making any changes
6. After I confirm, create missing accounts, add contacts, open opportunities
Multi-step outbound execution
Run outbound for [ACCOUNT NAME].
1. Pull their company record, existing contacts, and deal history from CRM
2. Find missing decision-makers via Apollo or LinkedIn
3. Build a growth strategy: which contacts, what angle, what timing
4. Draft personalised outreach emails for the top 3 contacts
5. Push new contacts to CRM
6. Show me the emails before anything is sent
Inbox triage and reply drafting
Triage my LinkedIn inbox.
1. List all unread conversations
2. Classify each by intent: positive, neutral, negative, needs follow-up
3. For every positive conversation, read the full thread and draft a reply that moves toward a call
4. Show me each draft before I approve
5. Send only after I approve each one individually
Proposal generation from transcript
Draft a proposal for [CLIENT NAME].
1. Pull the latest meeting transcript from Fireflies
2. Extract: their situation, stated problems, requirements, budget signals, timeline
3. Draft a full proposal: situation overview, proposed scope, deliverables, timeline, pricing
4. Flag anything unclear as a question rather than assuming
5. Show me the draft and iterate before publishing
Agentic safety features relevant to GTM teams
- Prompt injection resistance — better at recognising and refusing hijack attempts when running as an agent that browses the web or reads external content. Relevant for any workflow that pulls data from untrusted sources such as prospect websites or LinkedIn profiles.
- Malicious request refusal — cleaner and more consistent refusals. Doesn't stall or produce partial harmful output.
- Hallucination rate — lower than Sonnet 4.6. Important for any workflow that cites specific data, metrics, or company information.
- Sycophancy rate — lower than Sonnet 4.6. The model is less likely to agree with incorrect information in its context, which matters for review workflows where you want Claude to push back on errors.
Practical implication: you can give Sonnet 5 more autonomy than Sonnet 4.6. It's more likely to complete the task correctly and less likely to produce plausible-sounding wrong output.
Claude Code setup and skills for GTM engineers
Claude Code Desktop setup
Claude Code Desktop, not the CLI, is recommended for GTM engineers:
- Monitor all your sessions in one place
- Better UX for reviewing agent output before it sends
- Plugin support — connect Linear, CRM, and other tools directly
- Easy model switching between Sonnet 5 (execution) and Opus 4.8 (planning)
Setup:
- Download Claude Code Desktop
- Bottom-right dropdown: select Sonnet 5
- For planning sessions: switch to Opus 4.8 + Ultra Code or Max mode
- For execution: return to Sonnet 5 + medium effort
Installing skills for GTM work
Skills are SKILL.md files that live in ~/.claude/skills/. Sonnet 5 reads them before executing any workflow, and once installed they run with one slash command.
# Clone and install
git clone https://github.com/manojbajaj95/claude-gtm-plugin
cd claude-gtm-plugin && npm install -g @anthropic/claude-code
npx skills add manojbajaj95/claude-gtm-plugin
# Or install individual skills
npx skills add coreyhaines31/marketingskills --skill cold-email
npx skills add gtmagents/gtm-agents --skill cold-email-personalization
npx skills add sachacoldiq/ColdIQ-s-GTM-Skills
Build your CLAUDE.md before running any skill:
Build my CLAUDE.md. Ask me one question at a time about:
1. My ICP: exact titles, company types, sizes, buying signals
2. My offer in one sentence
3. My proof points with specific numbers
4. My voice: how I write, what I never say
5. My campaign history: what has worked
Save the output to CLAUDE.md in my project folder.
Once CLAUDE.md exists, every Sonnet 5 session reads it automatically. You never re-explain context again.
The Sonnet 5 + Opus 4.8 workflow in Claude Code
# Step 1: Planning (Opus 4.8, Ultra Code / Max mode, Plan mode)
"I want to build [DESCRIBE WHAT YOU WANT TO BUILD].
Plan this out fully. Ask me clarifying questions.
Build the architecture using sub-agents.
Save the plan to plan.md when done."
# Opus spins up Workflows (sub-agents) to design architecture
# Writes everything to a markdown plan file
# This is where you spend compute, so spend it well
# Step 2: Execution (Sonnet 5, Medium effort)
"Execute on the plan in plan.md."
# Sonnet 5 reads the plan and executes it
# No Opus tokens wasted on execution
# Same quality output, much lower cost
For ongoing GTM agent work with no new build needed, just use Sonnet 5 at medium effort:
/account-plan [client name]
/weekly-report
/draft-proposal [client name]
/crm-hygiene
/linkedin-outreach
/cold-email-sequence [persona]
Rate limits and availability
Sonnet 5 is the default model for Free and Pro plans as of June 30, 2026, and is also available on Max, Team, and Enterprise — in Claude Code and on the Claude Platform.
Rate limits have been increased across Chat, Cowork, Claude Code, and the Claude Platform to accommodate higher token usage from higher effort levels. There are three tiers: Start, Build, and Scale. Check your current tier at platform.claude.com/settings/limits.
Cyber safeguards are enabled by default on Sonnet 5 for cybersecurity tasks. If your team needs reduced cyber guardrails, use Opus 4.8 and enrol in the Cyber Verification Program.
Benchmarks, cost analysis, and the introductory pricing window
Benchmark data from the official Anthropic Sonnet 5 launch page. Use this for internal justification when switching your team to Sonnet 5.
Key benchmark results
Sonnet 5 vs Sonnet 4.6: substantially better on every benchmark tested — full improvement across reasoning, tool use, coding, and knowledge work.
Sonnet 5 vs Opus 4.8: doesn't beat Opus on any specific benchmark, but comes very close. Slightly better on knowledge work. On computer use and agentic search at high effort levels, it can match Opus 4.8.
BrowseComp (agentic search): Sonnet 5 at high effort approaches Opus 4.8 performance. Sonnet 4.6 fell well short of Opus 4.8 at any effort level.
OSWorld-Verified (computer use): same pattern. Sonnet 5 covers a much wider range of cost-performance options than Sonnet 4.6, and at high effort can match Opus 4.8 on some tasks.
Humanity's Last Exam: Sonnet 4.6 scores 34.6% without tools, 46.8% with tools (grader updated).
Safety: lower misaligned behaviour rate than Sonnet 4.6. Higher than Opus 4.8 and Mythos Preview.
Cybersecurity: Sonnet 5 was never able to develop a full working exploit in Firefox 147 vulnerability testing. It shows slightly higher partial success than Sonnet 4.6, due to general intelligence improvements rather than specific training.
Cost-performance comparison for GTM agents
For a similar agentic task at medium effort:
| Model | Cost | Notes |
|---|---|---|
| Opus 4.8 | $8 | Example cost from the YouTube breakdown |
| Sonnet 5 | ~$4 | At introductory pricing, roughly half |
Performance difference: roughly 5% lower pass rate on the same task.
For GTM teams spending $1,000–2,000/month on Claude tokens through Hermes or OpenClaw:
- Switching from Opus 4.8 to Sonnet 5 for execution work: roughly 50% cost reduction
- Performance impact: minimal for most GTM use cases — outreach, research, inbox, proposals
- Remaining Opus usage: planning, complex reasoning, tasks where quality is critical
Recommended split: 80% Sonnet 5 for execution and everyday agent work, 20% Opus 4.8 for planning and hard reasoning.
The introductory pricing window
Introductory pricing is $2/MTok input and $10/MTok output, available through August 31, 2026. Standard pricing from September 1, 2026 is $3/MTok input and $15/MTok output.
Sonnet 5 uses an updated tokenizer that can map the same input to 1.0–1.35x more tokens depending on content type. Introductory pricing is set to be roughly cost-neutral versus Sonnet 4.6 at these rates.
If you're currently paying for Sonnet 4.6 API usage, your bill should be roughly the same on Sonnet 5 through August 31, with meaningfully better output quality. After September 1, evaluate whether the quality improvement justifies the price difference for your specific workflows.
Quick reference: all commands and prompts
Switch to Sonnet 5 in your agent
# API model string
claude-sonnet-5
# Via agent (if not yet in model directory)
"Switch the Claude API model string to claude-sonnet-5. Look up the current string format and update the backend config."
# Settings config
{
"model": "claude-sonnet-5",
"effort": "medium"
}
Plan with Opus, execute with Sonnet 5
# PLANNING: Opus 4.8, Ultra Code, Plan mode
"Plan [WHAT YOU WANT TO BUILD]. Ask clarifying questions.
Use sub-agents to design the architecture. Save plan to plan.md."
# EXECUTION: Sonnet 5, Medium effort
"Execute on the plan in plan.md."
Build CLAUDE.md once
Build my CLAUDE.md. Ask me one question at a time about:
1. My ICP: exact titles, company types, sizes, buying signals, disqualifiers
2. My offer in one sentence
3. My proof points with specific numbers
4. My voice: how I write, what I never say
5. My campaign history: what has worked and what has not
Save to CLAUDE.md in my project folder.
Core GTM agent prompts
# Account planning
"Run account planning for [CLIENT NAME].
1. Pull company record, contacts, deal history from CRM
2. Find missing decision-makers via Apollo
3. Build growth strategy: which contacts, what angle, what timing
4. Draft personalised outreach for top 3 contacts
5. Push new contacts to CRM
6. Show emails before anything sends"
# Weekly sales report
"Run weekly report.
1. Pull full pipeline from CRM
2. Compare to last week's snapshot
3. Calculate changes: won, lost, stage movements, new opps
4. Pull latest team meeting transcript from Fireflies
5. Extract action items
6. Draft summary for Teams/Slack
7. Save new snapshot to CRM"
# Proposal from transcript
"Draft proposal for [CLIENT NAME].
1. Pull latest meeting transcript from Fireflies
2. Extract: situation, problems, requirements, budget signals, timeline
3. Draft: situation overview, scope, deliverables, timeline, pricing
4. Flag anything unclear as a question, do not assume
5. Show draft before publishing"
# CRM hygiene
"Run CRM hygiene.
1. Pull external meetings from last 2 weeks
2. Match attendee domains against CRM accounts using fuzzy logic
3. Flag attendees not in CRM, score by meeting relevance
4. Flag accounts with no open opportunity
5. Show report before making changes
6. After confirmation: create missing records, add contacts, open opps"
# LinkedIn inbox triage
"Triage LinkedIn inbox.
1. List all unread conversations
2. Classify each: positive / neutral / negative / needs follow-up
3. For each positive: read full thread, draft reply moving toward a call
4. Show each draft before I approve
5. Send only after individual approval"
# Cold email sequence
"Build a 6-touch cold email sequence for [PERSONA] at [COMPANY TYPE].
Trigger: [SIGNAL]. Offer: [ONE SENTENCE]. Proof: [SPECIFIC RESULT].
For each email: day, angle, 3 subject line variants, body.
Email 1: under 150 words. Emails 2-4: under 120. Email 5: under 80. Email 6 (breakup): under 60.
No em dashes. No circling back."
Install GTM skills
# GTM plugin (56 skills)
npx skills add manojbajaj95/claude-gtm-plugin
# Cold email skills
npx skills add coreyhaines31/marketingskills --skill cold-email
npx skills add gtmagents/gtm-agents --skill cold-email-personalization
# ColdIQ GTM skills (7 master skills, 52 sub-skills, 33 standalone)
npx skills add sachacoldiq/ColdIQ-s-GTM-Skills
# LinkedIn skills (sergebulaev, 10 skills)
npx skills add sergebulaev/linkedin-skills
# LinkedIn CLI (Linked API, 17 commands)
npm install -g @linkedapi/linkedin-cli && linkedin setup
# LinkedIn outreach skills (4 skills)
git clone https://github.com/claude-dev-code/claude-skills-linkedin
cp -R claude-skills-linkedin/linkedin-* ~/.claude/skills/
