Meetup #2

Claude CodeHelsinki

March 5, 2026

Built by all of us

A community of builders, creators, and developers exploring AI-assisted development together in the Nordics.

Your hosts

Hector Tortosa

Co-organizer

CTO & Co-founder at Nomain. Builder and Claude Community Ambassador.

Glafira Privalova

Co-organizer

iOS Software Engineer at F-Secure. Previously BeeHealthy. Metropolia & MSU alumni.

Stay connected

Websiteclaude-code-helsinki.codesharegrow.net

DiscordJoin the conversation

Community in numbers

Members

Meetups

Talks given

Speakers tonight

Tonight

194

signups

120

approved

44%

Daily users

23%

Regular users

22%

New & interested

11%

Occasional

29%

Engineers

22%

Founders

10%

Product & Design

Students

30%

Other

Top interests: Agents · Workflows · MCP · Multi-agent teams

Built with Opus 4.6

Claude Code Hackathon - Feb 10-16, 2026

13K+

applicants

500

selected

$100K

in API credits

Only 1 of 5 winners was a professional software engineer

Hackathon winners

1st

$50K

CrossBeamby Mike Brown

Speeds up California building permits - built by a personal injury attorney

2nd

$30K

Elisaby Jon McBee

Visual programming for kids - snap blocks, Claude builds the real code

3rd

$10K

postvisit.aiby Michal Nedoszytko

Turns visit transcripts into personalized health guidance - built by a cardiologist

Creative

$5K

Conductrby Asep Bagja Priandana

Play MIDI chords, Claude directs a generative band at ~15ms latency

Keep Thinking

$5K

TARAby Kyeyune Kazibwe

Dashcam footage to infrastructure investment recommendations - tested in Uganda

Claude Opus 4.6

February 5, 2026

Capabilities

1M token context window (beta)
128K max output tokens (2x vs 4.5)
Adaptive thinking replaces manual budget_tokens
Effort parameter GA: low, medium, high, max
Context compaction for infinite conversations
Fast mode: 2.5x faster at 6x price

Benchmark highlights vs 4.5

ARC AGI 2: 68.8% (+31.2)
BrowseComp: 84.0% (+16.2)
Humanity's Last Exam: 40.0% (+9.2)
OSWorld: 72.7% (+6.4)
Terminal-Bench 2.0: 65.4% (+5.6)
GPQA Diamond: 91.3% (+4.3)

Claude Sonnet 4.6

February 17, 2026

Capabilities

1M token context window (beta)
Within 1-2 pts of Opus on SWE-bench
Preferred over Opus 4.5 59% of the time
Effort parameter now on Sonnet too
Major prompt injection resistance gains
Adaptive thinking supported

What changed for users

Default model for Free and Pro plans
Free tier: file creation, connectors, skills
Same pricing as Sonnet 4.5
OSWorld: 72.5% (+11.1 vs 4.5)
SWE-bench: 79.6% (+2.4 vs 4.5)
+15 pts on heavy reasoning Q&A vs 4.5

Claude Code updates

v2.1.45 - v2.1.63 (Jan-Feb 2026)

releases

of GitHub commits

$2.5B

ARR

Agent Teams

research preview

A lead agent coordinates multiple teammates that work in parallel, each in its own context window and git worktree.

Shared task list with dependency tracking
Teammates send messages to each other
Each teammate has full tool access
Works on Bedrock, Vertex, and Foundry

// settings.json

{

"env": {

"CLAUDE_CODE_EXPERIMENTAL
_AGENT_TEAMS": "true"

}

/simplify & /batch

v2.1.63 - Boris Cherny: "I have been using both daily"

/simplify

Post-implementation cleanup before opening a PR

Spawns 3 parallel agents: reuse, quality, efficiency
Aggregates findings and applies fixes
/simplify focus on memory efficiency

/batch

Large-scale parallelizable code migrations

Decomposes work into 5-30 independent units
One agent per unit in isolated worktrees
/batch replace all uses of lodash with native

Memory & isolation

Auto-memory

Claude remembers what it learns across sessions

Build commands, test conventions, code style
Debugging solutions, architecture notes
Stored per-project in MEMORY.md
Shared across worktrees of same repo
Manage with /memory

Worktree isolation

Parallel sessions without conflicts

CLI: claude --worktree
Agent defs: isolation: worktree
Auto-cleanup if agent makes no changes
Background agents with background: true
Kill with Ctrl+F

More highlights

Remote Control - continue sessions from phone or browser, your local env stays available
HTTP hooks - POST JSON to URLs, not just shell commands
Context compaction - server-side summarization for effectively infinite conversations

Plugin marketplace - custom registries, version pinning
50+ memory leaks fixed - long sessions are stable now
~500ms faster startup - deferred imports, fewer HTTP calls
/voice - talk to Claude Code, fresh out of the oven

Platform updates

Feb 17

API tools GACode execution, memory, tool search, programmatic tool calling

Feb 17

Dynamic filteringMore accurate and token-efficient web search/fetch

Feb 19

Claude in PowerPointMCP connectors, 2x usage through March 19 (Pro plan)

Feb 19

Excel MCP connectorsS&P Global, LSEG, Daloopa, PitchBook, Moody's, FactSet

Feb 25

Cowork: Scheduled TasksRecurring automation, plugins for Jira, GitHub, Figma

Feb 25

Free Max Plan6 months free for open source maintainers

Trends

CLIs over MCPs

CLIs feel natural to coding agents in the terminal

Tools are shipping CLIs that agents can use directly, instead of (or alongside) MCP servers. CLIs are simpler to invoke, easier to test, and already how developers work.

Many now include install skills commands that add Claude Code skills automatically.

ObsidianCLI + skills for vault management

PlaywrightCLI with install-skills for browser testing

Hugging FaceCLI for model/dataset operations

Microsoft AspireCLI for .NET cloud-native apps

Rethinking CLAUDE.md

New research challenges common practices

The study

"Evaluating AGENTS.md" (arXiv 2602.11988) tested context files across 5,694 PRs from 12 repositories.

LLM-generated context files decrease performance by ~3%
All context files increase tool steps by 2-4 and costs by 20%+
Agents are highly compliant with instructions, but broader exploration rarely helps