What changed this week
On 30 June 2026, Anthropic shipped Claude Sonnet 5 and quietly made it the default across the Claude family — including inside Claude Code, the terminal-native agent that a large share of Indian and UK engineering teams now lean on for day-to-day work. The headline for builders is not a benchmark chart. It is three practical facts stacked together: Sonnet 5 is the new default, it carries a native 1M-token context window, and it is available at an introductory price of $2 per million input tokens and $10 per million output tokens through 31 August 2026, per Anthropic's announcement.
- New default — Sonnet 5 is the default for Free and Pro plans, available to Max, Team and Enterprise, and is the default model in Claude Code.
- 1M context, natively — the million-token window is the default and the maximum; there is no smaller variant to opt into.
- Introductory pricing — $2 in / $10 out per million tokens until 31 August, then standard rates of $3 in / $15 out.
- Most agentic Sonnet yet — Anthropic frames it as a substantial step up over Sonnet 4.6 on reasoning, tool use, coding and autonomous task handling.
If you have opened Claude Code since the start of the month and did not pin a model, you are almost certainly already running Sonnet 5. That is the quiet part worth saying out loud: the change reached your terminal before the changelog reached your inbox.
Check what you are actually running before you reason about cost. In Claude Code, confirm or set the model explicitly rather than trusting the default to stay put across updates:
# See which model the session is using
/model
# Pin Sonnet 5 explicitly for this project
/model claude-sonnet-5
# Or set it in the CLI for a one-off run
claude --model claude-sonnet-5 "refactor the billing module"Where Sonnet 5 sits in the Claude line-up
It helps to place this release inside a busy fortnight for Anthropic. On 1 July, Claude Fable 5 was redeployed globally after US export controls lifted on 30 June — that is the most capable generally available tier. Just before it, Claude Opus 4.8 introduced dynamic workflows and parallel subagents, aimed at the hardest planning and orchestration tasks. Sonnet 5 is the workhorse tier that sits below Opus: the model you reach for by default, not the one you save for the gnarliest problem.
That positioning matters for how you spend. Opus is where you pay a premium for peak capability on a minority of turns. Sonnet is where the volume lives — the file reads, the small edits, the test runs, the routine tool calls that make up the bulk of an agentic session. Making what Anthropic calls its strongest-yet Sonnet the default, and pricing it below its own standard rate for two months, resets the cost-per-task calculation for exactly the workloads that dominate a real engineering week.
The pricing picture, honestly
The introductory numbers are attractive, but read them with clear eyes. At $2 in / $10 out per million tokens, Sonnet 5 undercuts a lot of frontier-tier options for the next two months. After 31 August, though, the standard rate climbs to $3 in / $15 out — in plain arithmetic a 50% jump on both input and output once the promotion ends. That is a planning fact rather than a complaint: the launch price is a discount, not the resting price, so the number to build your budget around is the standard one. Treat the promo as a window to migrate and measure rather than a permanent floor, and if your spend is seasonal, note that the cliff lands just before many teams' Q4 planning cycle — model the September figure now so it does not distort a budget you set against August's invoice.
Here is how the current default and mid-tier options compare. All figures are list prices at the time of writing; verify against each provider's own pricing page before you commit spend, because these move often.
| Model | Price in / out (per MTok) | Context window | Agentic positioning |
|---|---|---|---|
| Claude Sonnet 5 (intro, to 31 Aug) | $2 / $10 | 1M tokens | Anthropic's most agentic Sonnet; default in Claude Code |
| Claude Sonnet 5 (standard, from 1 Sep) | $3 / $15 | 1M tokens | Same model, post-promo pricing |
| Google Gemini 3.5 Flash | ~$1.50 / $9.00 | 1M tokens | Budget-tier speed for high-volume work |
| OpenAI GPT-5.5 | $5 / $30 | 1M tokens | Frontier-tier general model; batch mode ~$2.50 / $15 |
The shape of the table is the story. All three flagships now offer a 1M-token window, so context length is no longer a differentiator you pay a premium for. The competition has moved to price-per-task and agentic quality. On raw list price Gemini 3.5 Flash is the cheapest and GPT-5.5 the dearest, with Sonnet 5's promo landing in between; Anthropic's pitch is that its coding and autonomous-agent quality justifies the position. For the economics of that trade-off in depth, our breakdown of Gemini 3.5 Flash cost-per-task and the GPT-5.5 API launch are useful companions.
List price is not your bill. Output tokens dominate agentic spend, and a chatty agent that writes long plans and verbose diffs can cost several times what a terse one does on the same task. The 31 August cliff turns a $10 output rate into $15 overnight — a 50% jump on the half of the equation that hurts most. Set a reminder now and re-run your cost model then.
What the 1M window actually unlocks
A native million-token context is not just a bigger bucket. For agentic coding it changes what fits in a single, coherent working session. A mid-sized service — say a payments module for a Bengaluru fintech, or a claims-processing backend for a London insurer — can now be read whole: source, tests, generated types and the relevant slice of documentation, all in one pass, without the lossy chunking that used to scatter cross-references. That is the same story we saw when Opus stepped up its long-context handling, now available at Sonnet economics.
The caveats from earlier long-context releases still apply. Reads at long context are reliable; autonomous writes across very large working sets can still drift, so keep the set of files an agent is actively editing modest and lean on retrieval for the wider background. And a million-token prompt priced at $2 per million input tokens is still a real number if you send it cold on every turn — which is exactly why caching is not optional at this scale.
There is a second, subtler benefit for teams working across time zones and hand-offs. A larger window means less state has to be reconstructed at the start of each session, which suits the reality of distributed teams — a Chennai engineer picking up where a London colleague left off can drop the whole relevant context into one prompt rather than stitching together fragments. Fewer chunk boundaries also means fewer places for the model to lose the thread on a long, multi-file change, which is precisely the kind of failure that used to force a manual clean-up pass.
Pair the 1M window with prompt caching. A stable system prompt plus a warm codebase context can bring the effective cost of a huge window close to a small one on repeated turns. Our guides on prompt caching for up to 90% savings and the broader cache, route and compress playbook cover the mechanics for Claude, GPT and Gemini alike.
What builders in India and the UK should do this week
The right move depends on how you consume the model. A few concrete steps, whether you are a two-person team in Chennai or an enterprise platform group in London running through AWS Mumbai or AWS London:
- Confirm your model. Run
/modelin Claude Code. If you are already on Sonnet 5 by default, your baseline just changed — re-benchmark your regular tasks rather than assuming last month's numbers hold. - Use the promo as a measurement window. Between now and 31 August, log cost-per-task on your real workloads at $2 / $10, then model the same tasks at $3 / $15 so the September step-up is not a surprise on the invoice.
- Route deliberately. Keep Opus 4.8 for the hardest planning and multi-subagent orchestration, default to Sonnet 5 for everyday agentic coding, and send bulk or trivial turns to a budget tier such as Gemini 3.5 Flash. Mixed routing beats a single default for most teams.
- Turn on caching before you scale. At 1M tokens the difference between warm and cold sessions is the difference between a viable unit economic and a scary one.
- Mind rate limits and metering. Anthropic raised rate limits alongside the launch, and if you are billing usage onward, our note on the Agent SDK metered-credits model is worth a read.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The bottom line
Claude Sonnet 5 is a workhorse upgrade dressed as a routine default swap. The million-token window, Anthropic's strongest-yet agentic Sonnet and a two-month introductory price make it the obvious baseline for Claude Code users right now. The discipline is in the details: budget for the standard rate, measure your own cost-per-task rather than trusting a launch chart, cache aggressively, and route Opus, Sonnet and a budget tier by job rather than by habit. Do that, and the reset in the cost-and-quality maths works in your favour — in Bengaluru, in London, and everywhere the terminal is open.
Primary details are in Anthropic's launch post at anthropic.com/news/claude-sonnet-5, with model specifics in the Claude Platform docs. Competitor list prices referenced from each provider's own pricing pages.